Words and rules : the ingredients of language

Steven Pinker

1

THE INFINITE LIBRARY

Language comes so naturally to us that it is easy to forget what a

strange and miraculous gift it is. All over the world members of our

species fashion their breath into hisses and hums and squeaks and pops and

listen to others do the same. We do this, of course, not only because we like

the sounds but because details of the sounds contain information about the

intentions of the person making them. We humans are fitted with a means of

sharing our ideas, in all their unfathomable vastness. When we listen to

speech, we can be led to think thoughts that have never been thought before

and that never would have occurred to us on our own. Behold, the bush

burned with fire, and the bush was not consumed. Man is born free, and

everywhere he is in chains. Emma Woodhouse, handsome, clever, and rich,

with a comfortable home and happy disposition, seemed to unite some of

the best blessings of existence. Energy equals mass times the speed of light

squared. I have found it impossible to carry the heavy burden of

responsibility and to discharge my duties as King without the help and

support of the woman I love. Language has fascinated people for thousands of years, and linguists

have studied every detail, from the number of languages spoken in New

Guinea to why we say razzle-dazzle instead of dazzle-razzle. Yet to me the

first and deepest challenge in understanding language is accounting for its

boundless expressive power. What is the trick behind our ability to fill one

another’s heads with so many different ideas?

The premise of this book is that there are two tricks, words and rules. They work by different principles, are learned and used in different ways,

and may even reside in different parts of the brain. Their border disputes

shape and reshape languages over centuries, and make language not only a

tool for communication but also a medium for wordplay and poetry and an

heirloom of endless fascination. The first trick, the word, is based on a memorized arbitrary pairing between

a sound and a meaning. “What’s in a name?” asks Juliet. “That which we

call a rose by any other name would smell as sweet.” What’s in a name is

that everyone in a language community tacitly agrees to use a particular

sound to convey a particular idea. Although the word rose does not smell

sweet or have thorns, we can use it to convey the idea of a rose because all

of us have learned, at our mother’s knee or in the playground, the same link

between a noise and a thought. Now any of us can convey the thought by

making the noise. The theory that words work by a conventional pairing of sound and

meaning is not banal or uncontroversial. In the earliest surviving debate on

linguistics, Plato has Hermogenes say, “Nothing has its name by nature, but

only by usage and custom.” Cratylus disagrees: “There is a correctness of

name existing by nature for everything: a name is not simply that which a

number of people jointly agree to call a thing.” Cratylus is a creationist, and

suggests that “a power greater than man assigned the first names to things.”

Today, those who see a correctness of names might attribute it instead to

onomatopoeia (words such as crash and oink that sound like what they

mean) or to sound symbolism (words such as sneer, cantankerous, and

mellifluous that naturally call to mind the things they mean).

Today this debate has been resolved in favor of Hermogenes’

conventional pairing. Early in this century Ferdinand de Saussure, a founder

of modern linguistics, called such pairing the arbitrary sign and made it a

cornerstone of the study of language.1 Onomatopoeia and sound symbolism certainly exist, but they are asterisks to the far more important principle of

the arbitrary sign—or else we would understand the words in every foreign

language instinctively, and never need a dictionary for our own!

Even the

most obviously onomatopoeic words—those for animal sounds—are

notoriously unpredictable, with pigs oinking boo-boo in Japan and dogs

barking gong-gong in Indonesia. Sound symbolism, for its part, was no

friend of the American woman in the throes of labor who overheard what

struck her as the most beautiful word in the English language and named

her newborn daughter Meconium, the medical term for fetal excrement.2

Though simple, the principle of the arbitrary sign is a powerful tool for

getting thoughts from head to head. Children begin to learn words before

their first birthday, and by their second they hoover them up at a rate of one

every two hours. By the time they enter school children command 13,000

words, and then the pace picks up, because new words rain down on them

from both speech and print. A typical high-school graduate knows about

60,000 words; a literate adult, perhaps twice that number.3 People

recognize words swiftly. The meaning of a spoken word is accessed by a

listener’s brain in about a fifth of a second, before the speaker has finished

pronouncing it. 4 The meaning of a printed word is registered even more quickly, in about an eighth of a second.5 People produce words almost as

rapidly: It takes the brain about a quarter of a second to find a word to name

an object, and about another quarter of a second to program the mouth and

tongue to pronounce it.6

The arbitrary sign works because a speaker and a listener can call on

identical entries in their mental dictionaries. The speaker has a thought,

makes a sound, and counts on the listener to hear the sound and recover that

thought. To depict an entry in the mental dictionary we need a way of

showing the entry itself, as well as its sound and meaning. The entry for a

word is simply its address in one’s memory, like the location of the

boldfaced entry for a word in a real dictionary. It’s convenient to use an

English letter sequence such as r-o-s-e to stand for the entry, as long as we

remember this is just a mnemonic tag that allows us to remember which

word the entry corresponds to; any symbol, such as 42759, would do just as

well. To depict the word’s sound, we can use a phonetic notation, such as

\[rōz].

* The meaning of a word is a link to an entry in the person’s mental

encyclopedia, which captures the person’s concept of a rose. For

convenience we can symbolize it with a picture, such as. So a mental

dictionary entry looks something like this:

rose

sound: rôz

meaning:

A final component is the word’s part of speech, or grammatical

category, which for rose is noun (N):

rose

sound: rōz

meaning:

part of speech: N

And that brings us to the second trick behind the vast expressive power of

language. People do not just blurt out isolated words but rather combine them into

phrases and sentences, in which the meaning of the combination can be

inferred from the meanings of the words and the way they are arranged. We

talk not merely of roses, but of the red rose, proud rose, sad rose of all my

days. We can express our feelings about bread and roses, guns and roses,

the War of the Roses, or days of wine and roses. We can say that lovely is

the rose, roses are red, or a rose is a rose is a rose. When we combine

words, their arrangement is crucial: Violets are red, roses are blue, though

containing all the ingredients of the familiar verse, means something very

different. We all know the difference between young women looking for

husbands and husbands looking for young women, and that looking women

husbands young for doesn’t mean anything at all. Inside everyone’s head there must be a code or protocol or set of rules

that specifies how words may be arranged into meaningful combinations. Modern linguists call it a grammar, sometimes a generative grammar to

distinguish it from the grammars used to teach foreign languages or to teach

the dos and don’ts of formal prose. A grammar assembles words into phrases according to the words’ part-

of-speech categories, such as noun and verb. To highlight a word’s category

and reduce visual clutter often it is convenient to omit the sound and

meaning and put the category label on top:

Similarly, the word a, an article or determiner, would look like this:

They can then be joined into the phrase a rose by a rule that joins a

determiner to a noun to yield a noun phrase (NP).

The rule can be shown as

a set of connected branches; this one says “a noun phrase may be composed

of a determiner followed by a noun”:

The symbols at the bottom of the branches are like slots into which words

may be plugged, as long as the words have the same labels growing out of

their tops. Here is the result, the phrase a rose:

With just two more rules we can build a complete toy grammar. One rule

defines a predicate or verb phrase (VP); the rule says that a verb phrase may

consist of a verb followed by its direct object, a noun phrase:

The other rule defines the sentence itself (S).

This rule says that a sentence

may be composed from a noun phrase (the subject) followed by a verb

phrase (the predicate):

When words are plugged into phrases according to these rules, and the

phrases are plugged into bigger phrases, we get a complete sentence, such

as A rose is a rose:

Other parts of the rules, not shown here, specify the meaning of the new

combination. For example, the complete NP rule says that the meaning of

the yellow rose of Texas is based on the meaning of rose, which is called the

head of the phrase, and that the other words modify the head in various

ways: yellow specifies a distinctive trait, Texas its location. These rules, though crude, illustrate the fantastic expressive power

made available by grammar. First, the rules are productive. By specifying a

string of kinds of words rather than a string of actual words, the rules allow

us to assemble new sentences on the fly and not regurgitate preassembled

clichés—and that allows us to convey unprecedented combinations of ideas. Though we often speak of roses being red, we could talk about violets being

red if the desire came over us (perhaps to announce a new hybrid), because

the rule allows us to insert violets into the N slot just as easily as roses. Second, the symbols contained by the rules are symbolic and hence

abstract. The rule doesn’t say, “A sentence may begin with a bunch of

words referring to a kind of flower”; rather, it says, “A sentence may begin

with an NP,” where NP is a symbol or variable that can be replaced by any

noun, just as x or y in a mathematical formula can be replaced by any

number. We can use the rules to talk about flowers and their colors and

smells, but we can just as easily use them to talk about karma or quarks or

floob-boober-bab-boober-bubs (who, according to Dr. Seuss, bounce in the

water like blubbery tubs).

Third, the rules are combinatorial. They don’t just have a single slot,

like a fill-in-the-blank exam question; every position in the sentence offers a

choice of words from a lengthy menu. Say everyday English has four

determiners ( a, any, one, and the) and ten thousand nouns. Then the rule for

a noun phrase allows four choices for the determiner, followed by ten

thousand choices for the head noun, yielding 4 x 10,000 = 40,000 ways to

utter a noun phrase. The rule for a sentence allows these forty thousand

subjects to be followed by any of four thousand verbs, providing 40,000 x

4,000 = 160,000,000 ways to utter the first three words of a sentence. Then

there are four choices for the determiner of the object (640 million four-

word beginnings) followed by ten thousand choices for the head noun of the

object, or 640,000,000 × 10,000 = 6,400,000,000,000 (6.4 trillion) five-

word sentences. Suppose it takes five seconds to produce one of these

sentences. To crank them all out, from The abandonment abased the abbey

and The abandonment abased the abbot, through The abandonment abased

the zoologist, all the way to The zoologist zoned the zoo, would take a

million years. Many such combinations are ungrammatical of course, owing to various

complications I haven’t mentioned—for example, you can’t say The Aaron,

a abandonment, or The abbot abase the abbey. And most of the

combinations are nonsensical: Abandonments can’t abbreviate, and abbeys

can’t abet. Yet even with these restrictions the expressive range of a

grammar is astonishing. The psychologist George Miller once

conservatively estimated that if speakers keep a sentence perfectly

grammatical and sensible as they choose their words, their menu at each

point offers an average of about ten choices (at some points there are many

more than ten choices; at others, only one or two).7 That works out to one

hundred thousand five-word sentences, one million six-word sentences, ten

million seven-word sentences, and so on. A sentence of twenty words is not

at all uncommon (the preceding sentence has twenty words before and so

on), and there are about one hundred million trillion of them in English. For

comparison, that is about a hundred times the number of seconds since the

birth of the universe. Grammar is an example of a combinatorial system, in which a small

inventory of elements can be assembled by rules into an immense set of

distinct objects. Combinatorial systems obey what Miller calls the

Exponential Principle: The number of possible combinations grows

exponentially (geometrically) with the size of the combination. 8

Combinatorial systems can generate inconceivably vast numbers of

products. Every kind of molecule in the universe is assembled from a

hundred-odd chemical elements; every protein building block and catalyst

in the living world is assembled from just twenty amino acids. Even when

the number of products is smaller, a combinatorial system can capture them

all and provide enormous savings in storage space. Eight bits define 28 =

256 distinct bytes, which is more than enough for all the numerals,

punctuation marks, and upper- and lowercase letters in our writing system. This allows computers to be built out of identical specks of silicon that can

be in just two states, instead of the dozens of pieces of type that once filled

typesetters’ cases. Billions of years ago life on Earth settled on a code in

which a string of three bases in a DNA molecule became the instruction for

selecting one amino acid when assembling a protein. There are four kinds

of bases, so a three-base string allows for 4 x 4 x 4 = 64 possibilities. That

is enough to give each of the twenty amino acids its own string, with plenty

left over for the start and stop instructions that begin and end the protein. Two bases would have been too few (4×4= 16), four more than needed

(4×4×4×4 = 256).

Perhaps the most vivid description of the staggering power of a

combinatorial system is in Jorge Luis Borges’s story “The Library of

Babel.”9 The library is a vast network of galleries with books composed of

all the combinations of twenty-two letters, the comma, the period, and the

space. Somewhere in the library is a book that contains the true history of

the future (including the story of your death), a book of prophecy that

vindicates the acts of every man in the universe, and a book containing the

clarification of the mysteries of humanity. People roamed the galleries in a

futile search for those texts from among the untold number of books with

false versions of each revelation, the millions of facsimiles of a given book

differing by a character, and, of course, the miles and miles of gibberish. The narrator notes that even when the human species goes extinct, the

library, that space of combinatorial possibilities, will endure: “illuminated,

solitary, infinite, perfectly motionless, equipped with precious volumes,

useless, incorruptible, secret.”

Technically, Borges needn’t have described the library as “infinite.” At

eighty characters a line, forty lines a page, and 410 pages a book, the

number of books is around 101,800,000, or 1 followed by 1.8 million

zeroes. That is, to be sure, a very large number—there are only 1070

particles in the visible universe—but it is a finite number. It is easy to make a toy grammar that is even more powerful than the

scheme that generates The Library of Babel. Suppose our rule for the verb

phrase is enriched to allow a sentence (S) to appear inside it, as in I told

Mary he was a fool, in which he was a fool comes after the object NP Mary:

Now our grammar is recursive: The rules create an entity that can

contain an example of itself. In this case, a Sentence contains a Verb Phrase

which in turn can contain a Sentence. An entity that contains an example of

itself can just as easily contain an example of itself that contains an example

of itself that contains an example of itself, and so on:

In this case a sentence can contain a verb phrase, which can contain a

sentence, which can contain a verb phrase, which can contain a sentence, ad

infinitum. For example, I think I’ll tell you that I just read a news story that

recounts that Stephen Brill reported that the press uncritically believed

Kenneth Starr’s announcement that Linda Tripp testified to him that Monica

Lewinsky told Tripp that Bill Clinton told Vernon Jordan to advise

Lewinsky not to testify to Starr that she had had a sexual relationship with

Clinton. That statement is a Russian doll with thirteen sentences inside

sentences inside sentences. A recursive grammar can generate sentences of

any length, and thus can generate an infinite number of sentences. So a

human being possessing a recursive grammar can express or understand an

infinite number of distinct thoughts, limited in practice only by stamina and

mortality. The idea that the creativity inherent in language can be explained by a

grammar of combinatorial rules is usually associated with the linguist

Noam Chomsky. Chomsky traced the idea to Wilhelm von Humboldt, a

nineteenth-century pioneer of linguistics, who explained language as “the

infinite use of finite media.” According to Chomsky, the idea is even older

than that; Humboldt was the last in a tradition of “Cartesian linguists”

dating back to the Enlightenment. 10

Enlightenment philosophers were captivated by the dizzying range of

thoughts made expressible by a combinatorial grammar. In his book The

Search for the Perfect Language the semiotician Umberto Eco recounts the

many Promethean schemes these philosophers came up with to perfect and

harness their power.11 Descartes noticed that the decimal system allows a

person to learn in a day the names of all the quantities to infinity, and he

suggested that a universal artificial language built on similar principles

could organize all human thoughts. Leibniz, too, dreamed of a universal

logical grammar that would generate only valid sequences of ideas,

banishing irrationality and error forever. Three hundred years later we still are fallible, and still take years to

learn a Babel of local languages with their tens of thousands of arbitrary

signs. Why has no modern language used the horsepower of combinatorial

grammar to the fullest and abandoned the unprincipled, parochial, onerous-

to-memorize laundry list called vocabulary?

The answer becomes clear

when we look at the most famous of the combinatorial schemes of the

Enlightenment, the philosophical language of Bishop John Wilkins. The

arbitrary name was an affront to Wilkins’s sense of good design, and he

strove for a way to eliminate it. He wrote, “We should, by learning...

the

Names of things, be instructed likewise in their Natures. ”

Wilkins’s system, laid out in a lengthy 1668 opus, offered the user a

non-arbitrary name for every thing by dividing the universe into categories

and subcategories and sub-subcategories, and assigning a vowel or

consonant to every branch in the tree. The first syllable identified one of the

forty categories into which Wilkins had sorted all thinkable thoughts. For

example, Z stood for “sensitive species” (animals) and could be followed

by i for “beasts” (quadrupeds).

The next consonant picked out a

subdivision; t, for example, stood for rapacious terrestrial European

canines. A final vowel pinpointed the species, yielding Zita as the name for

dogs. By similar computations one could deduce another two thousand

names for things. Zana is a scaly river fish with reddish flesh, in other

words, salmon. Sibα is a type of public military relation, namely, defense. Debα is a portion of the first of the terrestrial elements (fire), to wit, flame. Coba is a consanguinous economic relation of direct ascendant, a. k.a. father. Wilkins’s philosophical language has been analyzed insightfully by

Borges and Eco, and we can see why no one today speaks Wilkish.12 For

one thing, it forces users to perform a chain of computations in their heads

every time they want to refer to a dog. Every vowel and consonant is laden

with meaning and acts as a premise in a lengthening deduction. Speakers of

the language would have to play a game of Twenty Questions, inferring an

entity from a description, for every word in a sentence. They could of

course simply memorize the answers, such as that a portion of the first of

the terrestrial elements is a flame, but that is not much easier than

memorizing that the word for flame is flame. A second problem is that there are more things in heaven and earth than

were dreamt of in Wilkins’s philosophy, which identified only two thousand

concepts. Wilkins understood the exponential principle and tried to cope

with the problem by lengthening the words. He provided suffixes and

connectors that allowed calf, for example, to be expressed as cow + young,

and astronomer to be expressed as artist + star. But eventually he gave up

and resorted to using synonyms for concepts his language could not

generate, such as box for coffin. Wilkins’s dilemma was that he could either

expand his system to embrace all concepts, which would require even

longer and more unwieldy strings, or he could force his users to remember

the nearest synonym, reintroducing the despised memorization process. A third problem is that in a logical language words are assembled

purely on information-theoretic principles, with no regard to the problems

that incarnate creatures might have in pronouncing and understanding the

strings. A perfect combinatorial language is always in danger of generating

mouthfuls like mxyzptlk or bftsplk, so Wilkins and other language-designers

of the Enlightenment all had to make concessions to pronounceability and

euphony. Sometimes they defiled their systems with irregularities, for

example, reversing a vowel and consonant to make a word more

pronounceable. At other times they hobbled the system with restrictions,

such as that consonants and vowels must alternate. Every even-numbered

position in a word had to be filled by one of the nine vowels of English, and

that restricted many categories, such as species in a genus, to nine apiece,

regardless of how many species exist in the world. Another problem is that Wilkins’s words are packed tight with

information and lack the safety factor provided by redundancy. The

slightest slip of the tongue or pen guarantees misunderstanding. Eco catches

Wilkins himself misusing Gαde (barley) for Gαpe (tulip).

Finally, all that power is not being put to any sensible use. The beauty

of a combinatorial system is that it generates combinations that have never

before been considered but that one might want to talk about some day. For

example, the combinatorial system known as the periodic table of the

elements inspired chemists to look for hitherto unknown chemical elements

that should have occupied the empty slots in the table. Combinatorial

grammar allows us to talk about a combinatorial world, a world in which

violets could be red or a man could bite a dog. Yet familiar objects and

actions around us often form a non combinatorial list of distinctive kinds. When we merely have to single out one of them, a combinatorial system is

overkill. We never will have to refer to fish with an enmity to sheep or to

military actions with scales and reddish flesh, and that’s what a

combinatorial system for words like Wilkins’s allows us to do. To refer to

everyday things it’s easier to say dog or fish than to work through a

complicated taxonomy that is just a fancy way of singling out dogs or fish

anyway. The languages of Wilkins and other Enlightenment thinkers show that

combinatorial grammar has disadvantages as well as advantages, and that

illuminates our understanding of the design of human language. No

language works like Wilkins’s contraption, with every word compiled out

of meaningful vowels and consonants according to a master formula. All

languages force their speakers to memorize thousands of arbitrary words,

and now we can see why. 13 Many bodily organ systems are made from

several kinds of tissue optimized for jobs with contradictory specifications. Our eyes have rods for night vision and cones for day vision; our muscles

have slow-twitch fibers for sustained action and fast-twitch fibers for bursts

of speed. The human language system also appears to be built out of two

kinds of mental tissue. It has a lexicon of words, which refer to common

things such as people, places, objects, and actions, and which are handled

by a mechanism for storing and retrieving items in memory. And it has a

grammar of rules, which refer to novel relationships among things, and

which is handled by a mechanism for combining and analyzing sequences

of symbols. To a parsimonious scientific mind, however, two mental mechanisms

can be one too many. The poet William Empson wrote of the Latin

philosopher,

Lucretius could not credit centaurs;

Such bicycle he deemed asynchronous.14

Today’s skeptics also might wonder about a two-part design for language. Perhaps words and rules are two modes of operation of a single faculty. Simple, familiar thoughts need short noises, which we call words, and

complicated, unfamiliar thoughts need long noises, which we call phrases

and sentences. A single machine might make either short or long noises,

depending on the kinds of thoughts it is asked to express. Or perhaps there

is a gradual continuum between memory and combination rather than two

distinct mechanisms, with words at the memory end of the continuum and

sentences at the combination end. To show that words and rules are handled by different machines we

need to hold the input and output of the putative machines constant. We

need side-by-side specimens in which the same kind of thought is packed

into the same kind of verbiage, but one specimen shows the handiwork of a

word regurgitator and the other shows the handiwork of a rule amalgamator. I believe that languages do provide us with such specimens. They are called

regular and irregular words. English verbs come in two flavors. Regular verbs have past tense forms

that look like the verb with -ed on the end: Today I jog, yesterday I jogged. They are monotonously predictable: jog—jogged, walk—walked, play—

played, kiss—kissed, and so on. (Regular nouns, whose plurals end in -s,

such as cats and dogs, are similar.) The list of regular verbs is also open-

ended. There are thousands, perhaps tens of thousands, of regular verbs in

English (depending on how big a dictionary you consult), and new ones are

being added to the language all the time. When fax came into common

parlance a decade or so ago, no one had to inquire about its past-tense form;

everyone knew it was faxed. Similarly, when other words enter the language

such as spam (flood with E-mail), snarf (download a file), mung (damage

something), mosh (dance in roughhouse fashion), and Bork (challenge a

political nominee for partisan reasons), the past-tense forms do not need

separate introductions: We all deduce that they are spammed, snarfed,

munged, moshed, and Borked. Even young children do it. In 1958 the psychologist Jean Berko Gleason

tested four- to seven-year-old children with the following procedure, now

known as the wug-test:

This is a wug. Now there is another one. There are two of them. These are two \_\_\_\_\_\_\_\_.

The children could have refused to answer on the grounds that they had

never heard of a wug and had never been told how to talk about more than

one of them. Instead, Berko Gleason wrote, “Answers were willingly, and

often insistently, given.” Three-quarters of the preschoolers and 99 percent

of the first-graders filled in the blank with wugs. Similarly, when shown a

picture of a man who knows how to rick or bing or gling and did the same

thing yesterday, most children said that he ricked or binged or glinged. The children could not have heard their parents say wugs or binged

before entering the lab, because these words had been coined especially for

the experiment. Children therefore are not parrots who just play back what

they hear. And the children could not have been previously rewarded by

parents for uttering those forms, because the children did not know the

words before entering the lab. Children therefore are not like pigeons in a

Skinner box, who increase or decrease the frequency of responses in

reaction to the contingencies of reinforcement. Noam Chomsky and Eric

Lenneberg, pioneers of the modern study of language and contemporaries

of Berko Gleason in the Harvard-MIT community, pointed to children’s

ability to generalize constructions such as the regular past tense in support

of their theory that language is actively acquired by a special rule-forming

mechanism in the mind of the child. 15

As it happens, all children are subjects in a version of Berko Gleason’s

experiment. Children often make up words or mangle them and are happy

to put their new verbs in the past tense. Here are some examples:

spidered

lightninged

smunched

poonked

speeched

broomed

byed (went by)

eat lunched

cut-upped egg16

All children also make creative errors in their speech like these:

I buyed a fire dog for a grillion dollars. Hey, Horton heared a Who. My teacher holded the baby rabbits and we patted them. Daddy, I stealed some of the people out of the boat. Once upon a time a alligator was eating a dinosaur and the

dinosaur was eating the alligator and the dinosaur was eaten by

the alligator and the alligator goed kerplunk. 17

Such errors bring us to the second flavor of a verb in English: irregular. The past-tense form of an irregular verb is not simply the verb decorated

with an -ed ending. For example, the past tense of buy is not buyed, but bought. Similarly, the past tense of hear, hold, steal, and go are heard, held, stole, and went. Irregular verbs contrast with regular verbs in almost every way. Whereas regulars are orderly and predictable, irregulars are chaotic and

idiosyncratic. The past tense of sink is sank, and the past tense of ring is rang. But the past tense of cling is not clang, but clung. The past tense of think is neither thank nor thunk, but thought. And the past tense of blink is neither blank nor blunk nor blought, but a regular form, blinked. The language maven Richard Lederer wrote a poem, “Tense Times with Verbs,”

that begins:

The verbs in English are a fright. How can we learn to read and write?

Today we speak, but first we spoke;

Some faucets leak, but never loke. Today we write, but first we wrote;

We bite our tongues, but never bote. Each day I teach, for years I taught,

And preachers preach, but never praught. This tale I tell; this tale I told;

I smell the flowers, but never smold. If knights still slay, as once they slew,

Then do we play, as once we plew?

If I still do as once I did,

Then do cows moo, as they once mid?18

Also in contrast to the regulars, irregular verbs form a closed list. There

are only about 150 to 180 irregular verbs in modern English (depending on

how you count), and there have been no recent additions. 19 The youngest

irregular is probably snuck, which sneaked into the language over a century

ago and is still not accepted by purists. 20 And the freewheeling children in

Berko Gleason’s study were downright stodgy when it came to irregular

forms: Only one out of eighty-six turned bing into bang, and one other

turned gling into glang. 21

These differences suggest a simple theory. Regular past-tense forms are

predictable in sound and generated freely because they are products of a

rule that lives in the minds of children and adults: “The past tense of a verb

may be formed from the verb followed by the suffix -ed. ” The rule would

look just like the rules of syntax in the toy grammar we played with earlier,

and would generate a similar inverted-tree-like structure:

Irregular verbs, in contrast, are unpredictable in form and restricted to a

list because they are memorized and retrieved as individual words. An

irregular form would look just like the lexical entry we saw when

considering the name of the rose. It would be linked with the entry for the

plain form of the same verb and labeled as its past tense:

Two mechanisms trying to do the same job would get in each other’s

way unless something adjudicated between them, and there is indeed a

simple principle: If a word can provide its own past tense from memory, the

rule is blocked; elsewhere (by default), the rule applies. 22 The first part explains why we adults don’t say holded and stealed; our knowledge of

held and stole blocks the rule that would have added -ed. The second part

explains why both children and adults say Borked and moshed and ricked

and broomed; as long as a verb does not have a form in memory, the rule

may be applied. The ability of a rule to apply elsewhere or by default—that

is, to any word that does not already have a specified form in memory—is

the source of its power. A speaker who needs to express a past tense or

plural is never left speechless, even when a search in memory comes up

emptyhanded. The theory that regular forms are generated by rule and irregular forms

are retrieved by rote is pleasing not only because it explains the differences

in productivity between the two patterns but also because it fits nicely into

the larger picture of the design of language. At first glance irregular verbs would seem to have no reason to live. Why should language have forms that are just cussed exceptions to a rule?

What are they good for, besides giving children a way to make cute errors,

providing material for humorous verse, and making life miserable for

foreign language students?

In Woody Allen’s story “The Kugelmass

Episode” a humanities professor in a midlife crisis finds a magic cabinet

that projects him into any book he takes in with him. After a tempestuous

affair with Madame Bovary, Kugelmass tries again with another novel, but

this time the cabinet malfunctioned, and the professor “was projected into

an old textbook, Remedial Spanish, and was running for his life over a

barren, rocky terrain as the word tener (‘to have’)—a large and hairy irregular verb—raced after him on its spindly legs.” 23

But under the word-and-rule theory we need not suppose that evolution

fitted us with a special gadget for irregularity. Irregular forms are just

words. If our language faculty has a knack for memorizing words, it should

have no inhibitions about memorizing past-tense forms at the same time. These are the verbs we call irregular, and they are a mere 180 additions to a

mental lexicon that already numbers in the tens or hundreds of thousands. Irregular and regular forms therefore would be the inevitable outcome of

two mental subsystems, words and rules, trying to do the same thing,

namely, express an event or state that took place in the past. Regular and irregular forms throw a spotlight on the advantages and

disadvantages of words and rules, because everything else about them is the

same: They both are one word long, and both convey the same meaning,

past tense. The advantage of a rule is that a vast number of forms are

generated by a compact mechanism. In English the savings are significant:

The rules for - ed, -s, and -ing (the three regular forms of the verb) cut our

mental storage needs to a quarter of what they would be if each form had to

be stored separately. In other languages, such as Turkish, Bantu, and many

Native American languages, there can be hundreds, thousands, or even

millions of conjugated forms for every verb (for different combinations of

tense, person, number, gender, mood, case, and so on), and the savings are

indispensable. The rule also allows new words like mosh, rare words like

abase, and abstract words like abet to be supplied with a past tense

( moshed, abased, abetted), even if there were no previous opportunities for

the speaker and hearer to have committed the form to memory. On the other

hand, a rule is more powerful than needed for words we hear so often that

retrieval from memory is easy. As we shall see, it is the most common

verbs, such as be, have, do, go, and say, that turn out to be irregular in language after language. Rules have another shortcoming that invites the word system to

memorize irregulars. Recall that one of the nuisances plaguing John Wilkins

as he designed his perfect language was that flesh-and-blood humans had to

pronounce and understand the products of the rules. A sequence of sounds

that encodes a concept precisely and efficiently may be unresolvable by the

ear or unpronounceable by the tongue. So it is with the rule for the past tense in English. The delicate tongue-tap that graces the end of a regular

form may escape a listener and be omitted when he reproduces it, resulting

in a solecism such as suppose to, use to, or cut and dry, or in signs and inscriptions like these:

Broil Cod

Use Books

Whip Cream

Blacken redfish

Can Vegetables

Box sets

Handicap Facilities Available

In certain older expressions -ed was omitted so often that the expression

eventually lost the -ed altogether, even among careful speakers and

listeners. That’s how we ended up with ice cream (originally iced cream),

sour cream, mince meat, and Damn Yankees. 24 Irregular verbs, in contrast, tend to use vowel changes such as ring—rang, strike—struck, and blow—

blew, which are as clear as a bell. Similarly, the very obliviousness to the details of the verb that makes a

rule so powerful (it applies across the board to all verbs, whether they are

familiar sounding or not) can let it blindly jam a suffix onto the end of an

inhospitable sound. The result can be an uneuphonious tongue-twister such

as edited or sixths. Monstrosities like these are never found among the

irregulars, which all have standard Anglo-Saxon word sounds such as grew

and strode and clung, which please the ear and roll off the tongue.25

Language works by words and rules, each with strengths and

weaknesses. Irregular and regular verbs are contrasting specimens of words

and rules in action. These are the themes of this book, but with many twists

to come. It would be too good to be true if we reached a major conclusion

about the most complicated object in the known universe, the human brain,

simply by seeing how children name pictures of little birds. The word-and-

rule theory for regular and irregular verbs is an opening statement in the

latest round of a debate on how the mind works that has raged for centuries. It has inspired two alternative theories that are equally ingenious but

diametrically opposed, and intensive research showing what is right and

wrong about each of them—perhaps resolving the debate for good. The

theory has solved many puzzles about the English language, and has

illuminated the ways that children learn to talk, the forces that make

languages diverge and the forces that make them alike, the way that

language is processed in the brain, and even the nature of our concepts

about things and people. But to reach those conclusions we first must put

regular and irregular verbs under a more powerful magnifying glass, where

we will find some unexpected fingerprints. * This book uses a simplified phonetic notation similar to that found in dictionaries, in which the long vowels ā in bait, ē in beet, ī in bite, ō in boat, and ū in boot are distinguished from the short vowels ă in bat, ĕ in bet, ĭ in bit, ŏ in pot, and ŭ in but. An unadorned a stands for the first vowel in father or papa. The symbol is used for the neutral vowel in the suffix of melted and Hose’s (e. g., mĕlt d, rōz z), a version of the vowel sometimes called schwa. “Long vowel,” “short vowel,” and other technical terms in linguistics, psycholinguistics, and neuroscience are defined in the Glossary. 2

DISSECTION BY LINGUISTICS

Regular and irregular words have long served as metaphors for the law-

abiding and the quirky. Psychology textbooks point to children’s

errors like breaked and goed as evidence that we are a pattern-loving,

exception-hating species, explaining everything from why children have

trouble learning simple laws of physics to why adults make errors when

using computers or diagnosing diseases. In 1984 George Orwell has the

state banning irregular verbs as a sign of its determination to crush the

human spirit; in 1989 the writer of a personal ad in the New York Review of

Books asked, “Are you an irregular verb?” as a sign of her determination to

exalt it. Science is not always kind to folklore from the natural world. Elephants

do forget, lemmings don’t commit mass suicide, two snowflakes can be

alike, we use more than 5 percent of our brains, and Eskimos don’t have a

hundred words for snow. We had better give irregular and regular verbs a

closer look before using them as evidence for a language faculty that works

by words and rules, or more generally, a mind that works by lookup and

computation. Regular and irregular forms do not work in isolation; they are part of the

integrated living system we call a language. This chapter will tease out

regular inflection from the linguistic organs and tissues in which it is

embedded. The next chapter, on irregular verbs, will have a different feel. Living creatures can be dissected, but creatures dead so long that only a

trace of the living organs remain must be excavated. Our tour of the

irregular verbs will uncover them from layers of historical sediment laid

down over thousands of years. Does language even have an anatomy?

Many people think about

language in the following way: We need to communicate, and language is

the fulfillment of that need. For every idea there is a word and vice-versa,

and we utter the words in an order that reflects the connections among

ideas. If this common-sense view is true, there would be little need to speak

of language being a complex system. The complexity would reside in the

meanings, and language would reflect that complexity directly. The point of this chapter is to show that this view is mistaken. I will put

regular verbs under a microscope to reveal the delicate anatomy that makes

them work. Language does express meaning as sound, of course, but not in

a single step. Sentences are put together on an assembly line composed of

mental modules, shown on the following page. One is a storehouse of

memorized words, the mental lexicon. Another is a team of rules that

combine words and parts of words into bigger words, a component called

morphology. A third is a team of rules that combine words into phrases and

sentences, a component called syntax. The three components pass messages

about meaning back and forth with the rest of the mind so that the words

correspond to what the speaker wants to say. This interface between

language and mind is called semantics. Finally, the assembled words,

phrases, and sentences are massaged by a set of rules into a sound pattern

that we can pronounce when speaking or extract from the stream of noise

when listening. This interface between language and the mouth and ear is

called phonology. Many people are suspicious of box-and-arrow diagrams of the mind. The walls of the boxes and the paths of the arrows often seem arbitrary, and

could just as easily have been drawn differently. In the case of language,

however, these components pop out as we tease apart the phenomena, and

at least some of the divisions are now becoming visible in the living brain,

as we will see in chapter 9.

1 This chapter will explore the kinds of discoveries that have led linguists to divide language into parts, using only

the facts of regular and irregular words. First, we will see why the lexicon is

different from the two boxes of rules to the right, then why morphology is

in a different box from syntax, and finally, why phonology and semantics

each gets a box. The easiest boxes to keep separate ought to be the boxes containing words

and rules. From the discussion in the preceding chapter, it should be clear

that a simple word like duck belongs in the lexicon to the left in the

diagram. Just as clearly, a sentence like Daffy is a duck is assembled by the

rules of syntax in the box on the right. According to the words-and-rules

theory, irregular forms such as swam are also words that come from the

lexicon, because they are as arbitrary as duck. What do we do then with

regular forms like quacked?

They look like words and sound like words,

but I have been insisting they don’t have to be stored in the lexicon. They

don’t seem like words, but they don’t seem like sentences either, which are

the clearest products of rules. The problem is that the terms word and rule come from everyday

parlance and are as scientifically fuzzy as other vernacular terms, like bug

and rock. On closer examination, the word word has two very different

senses. 2 The first sense matches the everyday notion of a word: a stretch of

sound that expresses a concept, that is printed as a string of letters between

white spaces, and that may be combined with other words to form phrases

and sentences. Some of these words are stored whole in the lexicon, like

duck and swam; others are assembled out of smaller bits by rules of

morphology such as quacked and duck-billed platypus. A technical term for

a word in this sense is a morphological object, to be distinguished from

phrases and sentences, which are syntactic objects. The second sense of word is a stretch of sound that has to be memorized

because it cannot be generated by rules. Some memorized chunks are

smaller than a word in the first sense, such as prefixes like un- and re- and

suffixes like -able and -ed. Others are larger than a word in the first sense,

such as idioms, clichés, and collocations. Idioms are phrases whose

meanings cannot be computed out of their parts, such as eat your heart out

and beat around the bush. Collocations and clichés are strings of words that

are remembered as wholes and often used together, such as gone with the

wind or like two peas in a pod. People know tens of thousands of these

expressions; the linguist Ray Jackend-off refers to them as “the Wheel of

Fortune lexicon,” after the game show in which contestants guess a familiar

expression from a few fragments. A chunk of any size that has to be

memorized—prefix, suffix, whole word, idiom, collocation—is the second

sense of word. It is the sense of word that contrasts with rule, and the sense I had in mind when choosing the title of this book. A memorized chunk is

sometimes called a listeme, that is, an item that has to be memorized as part

of a list; one could argue that this book ought to have been called Listemes

and Rules. So walked is a word in the first sense (a morphological object) and not a

word in the second sense (a listeme); its listemes are walk and -ed. These

one-part listemes—prefixes, suffixes, and the stems they attach to, such as

walk—are called morphemes, a term coined by the nineteenth-century

linguist Baudouin de Courtenay to refer to “that part of a word which is

endowed with psychological autonomy and is for the very same reasons not

further divisible.” 3

What about the rules?

Why divide the rules of morphology, which build

complex words (including regular plurals and past-tense forms), from the

rules of syntax, which build phrases and sentences?

Both are productive,

recursive, combinatorial systems, and some linguists see them as two parts

of a larger system.4 Yet all linguists recognize that they are not identical. This may seem of no interest to anyone but a student cramming for a

Linguistics 101 final, but in fact it has been a source of countless barroom

arguments, late-night dorm-room debates, and irreconcilable differences. What is the correct word for people who pass by: passerbys or

passersby?

Do nervous fiancées dread the first meeting of the mother-in-

laws or the mothers-in-law?

Who did Richard Nixon force to resign: a

series of Attorney Generals, or a series of Attorneys General?

Here are a

few real-life examples:

Dear Ms. Grammar,

A member of the Friday Night Couples League...

had a hole in one

on the third hole and another on the fifth. Did he have two holes in

one or two hole in ones?

One of us believes that the pattern should

be the same as in attorneys general and passersby. The other

disagrees, believing that holes in one would indicate that the golfer

gained multiple holes in one shot. A Diet Coke has been wagered on

this, and we have agreed that Ms. Grammar shall be the final

authority. 5

SPOONFULS

From a recipe: “Now throw in two tablespoons full of chopped

parsley and cook ten minutes more. The quail ought to be tender by

then.” Never mind the quail; how are we ever going to get those

tablespoons tender?

The word, of course, is tablespoonfuls, no

matter how illogical it seems. One dictionary contains the entry

spoonsful, but this is not generally accepted. 6

Gin and tonic season (no hyphens, please) is just about finished, but

Joe Galeota of West Roxbury would still like to know how to order

when he’s having more than one. “Friends advised me that the

answer is gins and tonic’ because alcohol is the main ingredient,” he

writes. 7

Never has the U. S. faced a worse crisis than in 1887, after the

invention of the Jack-in-the-Box. It had become a fad overnight, and

everyone was having a whale of a time when someone asked, “What

is its plural?” “Jack-in-the-Boxes!” claimed some. Others hotly

insisted, “Jacks-in-the-Box!” Civil war seemed inevitable, when

Zeke Kelp’s Crusade won a compromise on “Jacks-in-the-Boxes.”

Unthanked for forty-three years, Kelp will be honored next week

when N. Y. City unveils a hydrant in his name.8

All right, the last example isn’t from real life; it’s from the Early

Cartoons and Writings of Dr. Seuss. The others are from well-known

language columnists. Hole-in-one is from Ms. Grammar, the nom de plume

of Barbara Walraff when presiding over “Word Court” in the Atlantic

Monthly. Spoonful is from Theodore Bernstein, the late New York Times

editor who wrote the syndicated column “Bernstein on Words.” Gin and

tonic is from Jan Freeman, who dispenses “The Word” in the Boston Globe. People disagree on how to pluralize nouns, and they care about who is

correct. Purists insist that the -s belongs on the noun in the middle of the

expression ( notaries public, runners-up), and those with the common touch

are content to leave it at the end ( notary publics, runner-ups).

“Ms. Grammar” advised her beseechers that holes in one is technically correct,

but added, “to say ‘two holes in one’ is to ask to be misunderstood.” Her

Solomonic suggestion was to say a hole in one twice, and to buy two Diet

Cokes. For my purpose—figuring out how the human mind deals with language

—there is no correct answer. Most disputes about “correct” usage are

questions of custom and authority rather than grammatical logic (see “The

Language Mavens” in my book The Language Instinct), and in these

disputes in particular, both parties have grammatical logic on their side. Their agony highlights the distinctions among lexicon, morphology, and

syntax, and illustrates the theme of this book: that the mind analyzes every

stretch of language as some mixture of memorized chunks and rule-

governed assemblies. How people pluralize an expression depends on how

they tacitly analyze it: as a word or as a phrase. With a simple word the plural suffix goes at the end: one girl, two girls. Now what happens in a compound word composed of two simple words,

such as cowgirl?

The plural still goes on the end: two cowgirls, not two cowsgirl or two cowsgirls. That is because the word girl inside cowgirl is special. It is called the head of the word, and it stands for the word as a

whole in determining its meaning (a cowgirl is a kind of girl) and in

determining its plural: The -s goes on girl. A phrase also has a head, and it

too determines the meaning and gets the plural. But now we discover the

major difference between a word, the product of morphology, and a phrase,

the product of syntax: In the phrase, the head is on the left, not the right. If

you meet more than one girl from Ipanema (head = girl), they are girls from

Ipanema, not girl from Ipanemas. With a word the plural is on the end

( cowgirls); with a phrase the plural can be in the middle ( girls from

Ipanema).9

The seeds of the mother-in-law dispute were sown by a special option of

English: Occasionally a phrase gets repackaged into a long word. For

example, a hangover victim may complain of a bottom-of-the-birdcage taste

in her mouth; the phrase bottom of the birdcage has been packaged as a

word that modifies taste. When a word-made-from-a-phrase is new and

fresh, speakers still can perceive the anatomy of the phrase inside the word. For example, we parse the modifier bottom-of-the-birdcage to understand

that it means something as foul as the bottom of a birdcage. But when the phrase is used as a word repeatedly, the original meaning

can recede from collective memory. The phrase boundaries melt into a glob,

and speakers no longer sense its parts. No one thinks of Thursday as Thor’s

Day anymore, or of breakfast as breaking a fast. Modern English has thousands of former phrases and complex words that have congealed into

what people now perceive as simple words, such as business (busyness),

Christmas (Christ’s Mass), and spinster (one who spins).

The meltdown, of

course, does not happen overnight or in all speakers at once; there must

have been a time when some English speakers still heard Christmas as

Christ’s Mass and others heard it as the arbitrary name of the holiday, just

as today’s older speakers hear the awe in awesome where younger speakers

hear the whole word as a synonym for good. Most of our disputed plurals originated as phrases and then became

words. Long ago people might have thought, “she is not my mother in

reality, she is only my mother in law” (that is, according to canon or Church

law).

But the concept of a spouse’s mother needs a word, and eventually the

phrase got reanalyzed as that word: “She is my mother-in-law. ” Similar

meltdowns occurred in these phrases:

Jack is in the box → That is a Jack-in-the-box. Phyllis completed that hole in one shot → She got a hole-in-one. Barry passed by → He is a passerby. I set aside a spoon full of parsley → I set aside a spoonful. If some speakers still hear the phrase inside the word, they will be tempted

to put the plural marker on the head of the phrase: two mother + s in law,

Jack + s in a box, hole + s in one, passer + s by, spoon + s full. But if speakers glom the words together in their minds, they will be tempted to put

the plural marker at the end: motherinlaw + s, jackinthebox + es, passerby

* s, holeinone + s, spoonful + s. It’s not that phrase hearers interpret these expressions literally (for

example, that a mother-in-law is a mother as recognized by the law), or that

the phrase-deaf treat them as any old string of consonants and vowels; both

surely recognize them as complex words built out of familiar words. It’s

just that they grow different kinds of connective tissue when piecing these

expressions together. Those who would describe themselves as sons-in-law

hear mother as the head of a phrase inside the word (shown in the left tree

in the diagram); those who would describe themselves as son-in-laws hear a

string of little words inside the big word (right tree):

A proof that the in-law expressions have congealed into words may be

found in the umbrella word in-law, which can stand alone and be pluralized

in the usual way: The in-laws are coming over. It is a good bet that many of

today’s commonly used phrases will also become opaque some day and turn

into words; the giveaway will be a plural at the end. Don’t be surprised if

one day you hear about grant-in-aids, bill of ladings, or work of arts. This ambiguity—one stretch of sound, two ways of building a tree in

the mind—also started the controversy raised by reports such as the

following:

While Mo Vaughn should finish well over.300 with close to 40

home runs and more than 100 RBIs, Mike Piazza has not been

producing anywhere close to what he did last season, when he hit.362 with 40 homers and 124 RBIs.10

Baseball purists who deplore artificial turf and the designated hitter get

equally incensed by the plural form RBIs. RBI is an acronym for run batted

in, a run scored by a teammate as a consequence of one’s batting the ball. An RBI and then another RBI are two runs batted in, and the acronym for

runs batted in is just RBI—so it should be 124 RBI, not 124 RBIs. (The purists are not mollified by the sportscasters’ common alternative, ribbies. )

But the purists fail to recognize that acronyms, like phrases, can turn into

bona fide words as a language evolves, as in TV, VCR, UFO, SOB, and PC. Once an acronym has become a word there is no reason not to treat it as a

word, including adding a plural suffix to it. Would anyone really talk about

three JP (justices of the peace), five POW (prisoners of war), or nine SOB

(sons of bitches)?

An additional puzzle surrounds governors-general, solicitors-general,

and attorneys-general. The speakers who bequeathed the plurals to us must

have analyzed the words as phrases, which have their heads on the left. Indeed, a governor-general is a general governor, namely, one who has

several governors under him. The puzzle is, why didn’t they simply call him

a general governor?

After all, the adjective comes before the head noun in

English, not after it. The answer is that these words, together with many

other terms related to government, were borrowed from French when

England was ruled by the Normans in the centuries after the invasion of

William the Conqueror in 1066.

In French, the adjective can come after the

head noun, as in États-Unis (United States) and chaise longue (long chair,

garbled into the English chaise lounge).

The earliest citation in the Oxford

English Dictionary is from 1292: “Tous attorneyz general purrount lever

fins et cirrographer” (All general attorneys may levy fines and make legal

documents).

Anyone who insists that we eternally analyze (hence pluralize)

these words as they were analyzed in the minds of the original speakers of

Norman French also should insist that we refer to more than one major

general as majors general, because a major-general was once a general

major (from the French major-général).

Long ago our linguistic foreparents

forgot the French connection and reanalyzed general from a modifying

adjective to a modified noun. So if you are ever challenged for saying attorney-generals, mother-in-

laws, passerbys, RBIs, or hole-in-ones, you can reply, “They are the very

model of the modern major general.” They come from reanalyzing a phrase

into a word, a common development in the history of English, and a nice

demonstration that we treat stretches of language not as sounds linked

directly to meanings but as structured trees. People who put different trees

on the same sound will use the sound in different ways, even if the meaning

is the same. Let’s now peer into the morphology box. Morphology may be divided into

derivation—rules that form a new word out of old words, like duckfeathers

and unkissable—and inflection—rules that modify a word to fit its role in a

sentence, what language teachers call conjugation and declension. The past

tense and plural forms are examples of inflection. English inflection is famous among linguists for being so boring. Other

languages exploit the combinatorial power of grammar to generate

impressive numbers of forms for each noun and verb. The verb in Spanish

or Italian comes in about fifty forms: first, second, and third persons, each

singular and plural, each in present, past, and future tenses, each in

indicative, subjunctive and conditional moods, plus some imperative,

participle, and infinitive forms. Languages outside the Indo-European

family, such as those spoken in Africa or the Americas, can be even more

prolific. In the Bantu language Kivunjo, for example, a verb is encrusted

with prefixes and suffixes that multiply out to half a million combinations

per verb. 11 But English speakers subsist on only four:

open

opens

opened

opening

Strangely enough, English grammar does not have only four roles for verbs

to play. It has at least thirteen different roles, but it shares the four forms

among them, as if suffixes were expensive and the designers of the

language wanted to economize. The first suffix is a silent bit of nothing, -Ø, which when added to the

stem open turns it into the inflected form open. You may wonder: Why say

that speakers hallucinate an imaginary suffix at the end of a word?

The

reason is that it distinguishes the root or stem—the irreducible nugget found

in the mental dictionary that captures the essence of a verb and upon which

suffixes are hung—from a particular incarnation of that verb with a

particular person, number, and tense. In English they can sound the same—

to open and I open—which disguises the fact that they are different versions

of the verb. In other languages the form of the verb that you look up in a

dictionary cannot be pronounced. For example, in Spanish you can say

canto, cantéis, canten, and so on, leaving cant- as the stem, but you can never say cant- by itself. Stems are therefore not the same things as

pronounceable verb forms, and that distinction is useful to preserve in

English— to open versus open Ø—even though the two forms sometimes

sound the same. The suffix, -Ø is used in four variations of the verb in English:

Present tense, all but third-person singular: I, you, we, they open it. Infinitive: They may open it, They tried to open it. Imperative: Open!

Subjunctive: They insisted that it open. The suffix -s is used for only one purpose:

Present tense, third-person singular: He, she, it opens the door. The suffix -ing is used in at least four ways:

Progressive participle: He is opening it. Present participle: He tried opening the door. Verbal noun (gerund): His incessant opening of the boxes. Verbal adjective: A quietly-opening door. Finally we come to our friend -ed, which has four jobs:

Past tense: It opened. Perfect Participle: It has opened. Passive Participle: It was being opened. Verbal adjective: A recently -opened box. 12

Why make all these distinctions among verb forms that sound the same?

One reason is that the list of phrases calling for a form such as opened have

nothing in common: To capture the behavior of -ed, we have no choice but

to list four phrase types separately. Another reason is that some distinctions

that are inaudible for regular verbs are audible for irregular ones, and this

shows that English speakers register these distinctions as they speak. About

a third of the irregular verbs have different forms for the stem, the past

tense, and the perfect participle: I sing, I sang, I have sung; I eat, I ate, I

have eaten. A few make a further distinction and have a special form for the

verbal adjective —a newly wedded couple; a drunken sailor; a shrunken

head; rotten eggs—which is not used for the participle: people say They

have wed, not wedded; He has drunk, not drunken; It has shrunk, not shrunken; The eggs have rotted, not rotten. And one verb comes in eight different forms:

Infinitive; subjunctive; imperative: To be or not to be; Let it be; Be

prepared. Present tense, first-person singular: I am the walrus. Present tense, second-person singular, all persons plural:

You/we/they are family. Present tense, third-person singular: He/she/it is the rock. Past tense, first- and third-person singular: I/he/she/it was born by

the river. Past tense, second-person singular, all persons plural; subjunctive:

The way we/you/they were; If I were a rich man. Progressive and present participle; gerund: You’re being silly; It’s

not easy being green; Being and Nothingness. Perfect participle: I’ve been a puppet, a pauper, a pirate, a poet, a

pawn and a king. With nouns, too, different grammatical forms have to dip into the same

small pool of suffixes. The naked stem dog must be distinguished from the

singular dog + Ø because a dogcatcher doesn’t catch just one dog and a dog

lover doesn’t love just one. The dog inside these compounds refers to dogs

in general and thus differs in meaning from the singular form in a dog. The

plural dogs uses -s, which we have already met in the verb system in She

opens the door. The possessive forms dog’s (singular) and dogs’ (plural) use

it too; the three noun forms dogs, dog’s, and dogs’ differ only in

punctuation. All this redundancy suggests that regular inflection in English is

remarkably simple. All the inflections are suffixes; none of the grammatical

roles call for a prefix or some other way of decorating or tinkering with a

word. And every word has at most one inflectional suffix. We never get

opensed or opensing, nor do the plural -s and possessive s stack up when several owners own something: the dogs’ blanket, not the dogs’s (dogzez)

blanket. Finally, each niblet of sound making up a suffix has a life of its

own and combines with several verb forms, noun forms, or both, rather than

being a slave to only one role. This suggests that instead of crediting

English speakers with seventeen verbose rules like “To form the past tense,

add -ed to the end of the verb,” we can credit them with just one rule:13 “A word may be composed of a stem followed by a suffix,” like the simple rule

shown on page 16.

All the other details can be handled by assuming that

suffixes are stored in the mental lexicon with entries like those for words,

perhaps something like this:

-ed

sound: d

part of speech: suffix

use 1: past tense of a verb

use 2: perfect participle of a verb

use 3: passive participle of a verb

use 4: adjective formed from a verb

By factoring seventeen verbose rules into one austere rule and four

lexical entries, one per suffix, we not only save ink but get some insight into

the mental organization of language. English could have used seventeen

different forms for its seventeen slots in the noun declension and verb

conjugation: prefixes such as ib-, tra-, and ka-, suffixes such as -og, -ig, and

-ab, and so on. Instead the slots share a few sounds (-0, -ed, -s, -ing) and

one position (immediately following the verb).

This miserliness, called

syncretism, is found in language after language. Syncretism suggests that

the mind keeps separate accounts for the templates that build words (for

example, “word = stem + suffix”), for the scraps of sound that may be

added to words ( -s, -ed, and -ing), and for the roles these additions can play (for example, plural, participle, imperative).

14 A particular construction like the English past tense is a mix-and-match affair, assembled by hooking

together parts also used in other constructions. No one knows why

languages like to recycle their suffixes and other ways of modifying words. It’s certainly not to save memory space, because the savings are trivial. Perhaps the reason is to help listeners recognize when a word is composed

of a stem and a suffix rather than being a simple stem. Whatever its

purpose, syncretism shows that in the language system, combination is in

the blood; even the tiniest suffixes are combinations of smaller parts. Syncretism—one form, several roles—is one kind of violation of the

simplest conceivable system in which every sound has one meaning and

vice-versa. The other kind of violation—one role, several forms—is

rampant in languages as well; linguists call it allomorphy.15 Take the

regular past-tense suffix—or is it suffixes?

Though always spelled -ed, it is

pronounced in three different ways. In walked, it is pronounced t. In jogged,

it is pronounced d. And in patted, it is pronounced d, where is a neutral

vowel called “schwa.” We also find allomorphy in the regular plural: The

suffix -s has three different forms in cats, dogs, and horses. Are there in fact three past-tense suffixes and three plural suffixes?

In

some languages, we are forced to this messy conclusion. Dutch speakers,

for example, select either -en or -s as the regular plural, depending on the

sound of the end of the noun. But in English the three-way variation has a

simpler explanation, worked out by the linguists Arnold Zwicky and Alan

Prince. One past tense suffix is stored in the lexicon, not three, and a

separate module fiddles with its pronunciation: the rules of phonology,

which define the sound pattern or accent of a language. 16

Why do we pronounce the past tense suffix as t in walked, d in jogged,

and id in patted?

The choice is completely predictable, and can be stated as

a list of rules:

1.

Use id if the verb ends in t or d (for example, in patted and padded).

2.

If it doesn’t, use t if the verb ends in an unvoiced consonant—

that is, a consonant in which the vocal cords don’t buzz, namely

p, k, f, s, sh, ch, and th (for example, tapped, walked, sniffed,

passed, bashed, touched, and frothed).

3.

Use d for all other verbs: those ending in vowels, such as played

and glowed, and those ending in the voiced consonants l, r, m, n,

b, g, v, z, j, zh, and th (for example, smelled, marred, slammed,

planned, scrubbed, pegged, saved, buzzed, urged, camouflaged,

and bathed).

This sounds like something out of the tax code. Let’s see if we can do

better. The first thing to notice is that nothing in these rules is specific to the

past tense. Other constructions that use -ed work the same way:

Outside the verb system entirely is yet another -ed construction that comes

in the three variations; it turns a noun that means “X” into an adjective that

means “having X”:

The regular plural -s also comes in three forms, which you can hear in

hawks, dogs, and horses. The variation mirrors the past tense uncannily. Use z. when the noun ends in a sibilant sound: s, z, sh, zh, j, or ch. If it doesn’t, use s if the noun ends in an unvoiced consonant. Use z for all other

nouns. In fact, not only does this pattern appear with the plural, it appears

with the other -s suffixes as well:

The variation even appears in versions of -s that aren’t genuine suffixes. English speakers commonly contract the verbs has, is, and does to their

final consonant and glue it onto the end of the subject, as in Mom’s left or

Dad’s home. Sure enough, the contraction is pronounced in three ways,

depending on how the noun ends:

That’s not all. English has an affective -s that can be used to form

nicknames in some dialects and argots, as in Pops, Moms, Fats, Pats, and

Wills (the prince second in line to the British throne).

That -s can also show

up in emotionally colored slang such as honkers and nuts, similar to the -y

and - o that give us batty and wacko. (Sometimes the two suffixes are even

used together, as in Patsy, Bugsy, Mugsy, footsie, fatso, and Ratso. ) Still another version of -s appears in adverbial forms such as unawares,

nowadays, besides, backwards, thereabouts, and amidships. A final use for

s is as a meaningless link joining the words in compounds such as

huntsman, statesman, kinsman, bondsman, Scotsman, and grantsmanship. And yes, all of these -s’s can be pronounced either as s or as z, depending

on the preceding consonant (it’s hard to come up with examples for the

third column):

So we have fifteen suffixes that show the same three-way or two-way

variation. Forty-one suffixes that happen to fall into fifteen parallel sets of

alternatives is too much of a coincidence to stomach. More likely, one set of

rules creates the three-way variation, and the set applies in at least fifteen

situations. There is a second, equally striking set of coincidences that runs across

the suffixes. If the variation came from any old set of if...

then rules, we

would expect to find all kinds of pairings between stems and suffixes: for

example, “Use s after the vowels a and e or after the consonants th and g, ”

“Use d after a k,” and so on. But the rules are far more lawful than that. The

t sound comes after unvoiced consonants, and the t itself is unvoiced. The d

sound comes after voiced sounds, and the d itself is voiced. The -s suffixes

show the same chameleonlike behavior: We find unvoiced s after unvoiced

consonants, and voiced z after voiced consonants. It looks as if something is

trying to keep the consonants at the end of a word consistent: All of them

are voiced, or all of them are unvoiced. Indeed, something is—the sound pattern of the English language. English never forces speakers to turn their vocal cords on for one consonant

then off for the next, or vice-versa. We see the restriction in force in one-

piece words that end in a cluster of consonants. These words never received

a suffix; they just happen to be built that way, so any sound pattern they

display cannot have come from a suffix rule, but rather from the way

English speakers like to pronounce words in general. In all but one of these

words, the vocal cord switch can be left in the “off” position:

In one English word, adze, the vocal cord switch is left in the “on” position:

In no English word is the voicing switch toggled on and off, in an ending

like zt, gs, kz, or sd. These difficult-to-pronounce clusters can, however, be created by a

dumb rule of morphology that pins a suffix onto the end of a word without

regard for how the resulting train of consonants is to be pronounced. That is

what happens when a rule adds a d sound to walk or an s sound to dog. English cleans up these awkward mismatches with a different kind of rule. The rule says, “When there is a cluster of consonants at the end of a

syllable, adjust the voicing setting of the last consonant to make it

consistent with its neighbor on the left.” (In other words, change kz to ks, pd

to pt, and so on.) The rule does not care whether the syllable was formed by

a past-tense suffix, a plural suffix, a contracted has, a nickname with -s, or

anything else. It kicks in after the syllable has been assembled, in the

cleanup module we call phonology. Can we now tell whether the suffix stored in the lexicon is -d, and is

converted to a t when it finds itself at the end of walk, or whether it is -t and

is converted to d when it finds itself at the end of jog?

A little detective work can settle the question. Not every sound cares about the consonant

that follows it. Those that do are consonants in which the airstream is

obstructed, namely p, h, t, d, k, g, s, sh, ch, z, zh, and th. But the vowels, and

the vowel-like consonants r, l, n, and m, are indifferent to what comes after

them; they tolerate either s or z, either t or d, as we see in these one-piece words:

Here we have laissez-faire environments in which the suffixes can show

their true colors, untouched by rules of phonology. What do we find?

That

the virgin suffixes are pronounced -d and -z, not -t and -s:

The -t and -s we hear in words with choosy sounds such as walked and cats

must be the aftermath of the rule. Finally, what about the funny extra vowel in patted and horses?

Here

again the change in sound is not some random act of vandalism. The vowel

appears when d follows t or d, and when z follows s or z. The word endings that trigger the extra vowel are similar in pronunciation to the suffixes

themselves, and that can’t be a coincidence. Apparently a rule is trying to

separate too-similar adjacent consonants by pushing a vowel between them:

between t and d, d and d, s and z, z and z, sh and z, and so on. In many languages the rules of phonology do something when a rule of morphology

leaves two identical or near-identical consonants in a row, presumably

because there’s no natural way to pronounce them. Some languages drop

the second consonant, others merge the two into one long consonant, and

still others, like English, wedge a vowel between them. As with the rule that

fiddles with voicing, the rule that inserts a vowel must live in a phonology

module separate from rules that stick on the various suffixes, because the

rule is oblivious to what kind of suffix it manipulates. We even can deduce which of the two rules applies first, the one that

changes the voicing setting or the one that inserts the vowel. The devoicing

rule is triggered by adjacent consonants; the vowel rule breaks up adjacent

consonants. If the voicing rule came first, it would convert pat + d to pat +

t, and only then would the vowel be inserted, yielding păt t:

But that is not how we pronounce it; we say păt d. This means that the

vowel rule must have come first, creating patted; now the voicing rule is no

longer compelled to do anything, because the td sequence that would trigger

it has been broken up:

The ordering makes sense when you think about how the phonology

module should be organized. It has some rules that edit the string of vowels

and consonants composing a word (phonology proper), and other rules that

convert the string into actual sounds or muscle movements (phonetics).

The

vowel-insertion rule makes a major change in the stuff that makes up a

word, and belongs in the first subcomponent; the voicing rule does a last-

minute adjustment of pronunciation for the benefit of the muscles, and

belongs in the second.17

This completes the analysis of the three versions of the past-tense suffix. When we started, we needed forty-odd rules, each stipulating that some

suffix be placed next to some word ending. We have ended up with just two

rules. Best of all, what the rules do, why they do it, and in what order they

do it all make sense in the light of the sound pattern of English. Indeed, this

kind of layering may be found in languages all over the world. Incidentally, there is corroborating evidence of a completely different

kind that shows that the three forms of -ed and -s are created on the fly by a

phonological rule. Some psycholinguists keep a pad and pencil in their

pockets and write down every slip of the tongue they hear. People make one

or two such errors for every thousand words they say, and many of the

errors consist in deleting, repeating, or switching around vowels or

consonants. 18 The last kind oí error is called a Spoonerism, in honor of the

Reverend William Spooner (1844-1930), warden of New College at Oxford,

who came out with surprises such as Our queer old dean, You have hissed

all my mystery lessons and tasted the whole worm, and It is now kistomary

to cuss the bride. They sound too good to be true, but I have heard similar

errors myself. After I spoke at a scientific symposium the chair wrapped up

the session by saying I would like to spank the speakers, and when I asked a

friend how he liked his new condominium, he said It seats my nudes. Speech errors provide clues on how the speech system is organized. For

example, when a person intends to say grapefruits but accidentally leaves

out the t, how does he pronounce the plural?

If there were a distinct plural

suffix pronounced -ss, he would say grapefrooss, since this is what the t in

the grapefruit entry would have demanded. In fact he says grapefrooz—

pronouncing the plural as z, which is appropriate to words ending in a

vowel. 19 Similarly, a person may say The infant tucks—touches the nipple, not tuck- z, or may say Did you buy enough breakfas z?

, not breakfass. The

errors show that the form of the suffix must be computed after the vowels

and consonants of the noun or verb were placed on the chute to the vocal

tract. English did not always have single-consonant suffixes and a rule that

separates them from a too-similar word ending. Our current system is the

result of a reorganization that began around the time of the origin of

Modern English in the seventeenth century. Before that, -ed and -s suffixes

were pronounced (and spelled) with vowels all the time, not just with words

ending in t or d or in s or z. For centuries, English speakers had been concentrating stress on the first syllables of words, which shriveled the later

syllables, and speakers began to leave out the vowels in the suffixes of

many words. Writers called attention to the new, clipped pronunciations by

spelling them phonetically with an apostrophe in place of the deleted vowel,

as in Shakespeare’s play about “a pair of star-cross’d lovers”:

Death, that has suck’d the honey of thy breath,

Hath no power yet upon thy beauty:

Thou art not conquer’d; beauty’s ensign yet

Is crimson in thy lips and in thy cheeks. The guardians of the English language deplored the change, as they do

all changes. In “A Proposal for Correcting, Improving, and Ascertaining the

English Tongue,” Jonathan Swift wrote:

What does your lordship think of the words “drudg’d,” “disturb’d,”

“rebuk’d,” “fledg’d,” and a thousand others everywhere to be met

with in prose as well as verse?

Where, by leaving out a vowel to

save a syllable, we form so jarring a sound, and so difficult to utter,

that I have often wondered how it could ever obtain. His contemporary, Samuel Johnson, who was standardizing the spellings of

English words in a way that reflected the morphemes that composed them,

recognized that ’d and -ed were the same morpheme, and obliterated the

distinction in their spelling, making ed the spelling for both. 20 It is unclear why he chose to leave the e in -ed across the board ( mapped and matted), but opted to spell -s either with or without an e, depending on how it is pronounced ( maps and masses).

Today the old syllabic suffix survives in a handful of adjectives:

accursed, aged, beloved, bended (in the expression on bended knees),

blessed, crooked, cussed, dogged, jagged, learned, naked, ragged, wicked,

and wretched. (A few more survive in rural dialects, such as fork è d, peakèd,

streakèd, and strip è d.)21 Many of them are archaic or poetic and are used mainly in self-conscious speech. The psychologist Melissa Bowerman, a

researcher of child language, had this exchange with her four-year-old

daughter about a class trip to a natural history museum:22

MOTHER ( playfully): Maybe you’ll see something wingèd. DAUGHTER: Maybe we’ll see something snakèd!

We’ve seen why the syntax box, which builds phrases and sentences, has to

be separated from the morphology box, which builds words. We also have

seen why the phonology box, which massages words into a pronounceable

stream of sound, has to be separated from syntax, morphology, and the

lexicon. But why do we need separate boxes for semantics (the thoughts

expressed in language) and the lexicon?

Could we reduce the difference

between regular and irregular verbs to a difference in meaning between the

two kinds of verbs, rather than putting one kind in the morphology box and

the other in the lexicon?

Do we even need to talk about an “entry in the

mental lexicon,” the address in memory that holds a link to a sound and a

link to a meaning?

Or could we connect thoughts to sounds directly,

eliminating the middleman?

Here are some facts that suggest that we do

need to credit the human mind with something like dictionary entries. First, the English irregular verbs could not have arisen simply from a

communal effort to optimize clarity. While irregular forms on average are

harder to mistake for their base forms than regular forms are ( bring doesn’t

sound like brought, nor take like took), many irregulars are identical to their base forms: Today I hit, yesterday I hit; Today I put, yesterday I put. A

sentence such as On Wednesday I cut the grass could mean last Wednesday,

next Wednesday, or every Wednesday. If cut were regular, the ambiguity

would never arise: On Wednesday I cutted the grass would single out the

preceding Wednesday. Despite the potential ambiguity, however, twenty-

eight English verbs insist on remaining unchanged in the past tense. Also, irregular forms do not correlate with any kind of meaning. Many

verbs are similar in meaning but have completely different past-tense forms. For example, hit, strike, and slap all refer to hitting. Hit is an irregular verb that does not change in the past tense: Today we hit golf balls; Yesterday we

hit golf balls. Strike is an irregular verb that changes its vowel, yielding

struck. And slap is a regular verb, with past tense slapped. Not only are there verbs with similar meanings and different past-tense

forms, there are verbs with different meanings and the same past-tense

forms. English has a class of verbs linguists call light verbs, such as come, go, do, take, have, set, get, put, and stand. Compared to ordinary verbs they

are less filling; a light verb doesn’t have a meaning that stays with it, but

takes on dozens of meanings, especially in combination with particles such

as in, out, up, off, over, and around:

come (move to here), come around (agree), come into (inherit),

come (reach orgasm), come off as (appear), come out (divulge

homosexuality), come to (awaken)

go (move to there), go out with (date), go nuts (dement), go in for

(choose), go off (explode), go off (spoil)

do (act), do in (kill), do up (decorate), do a number on (overwhelm),

do lunch (eat together)

take (cause to go with), take in (swindle), take off (launch), take in

(welcome), take over (usurp), take up (commence), take a leak

(urinate), take a bath (lose money), take a bath (bathe), take a

walk (walk), take a look (look)

have (possess), have (eat), have (seduce), have a heart (sympathize),

have aver (entertain), have a cow (be angry)

get (retrieve), get (become), get over (survive), get out (divulge), get off on (enjoy), get a life (self-improve)

set (place), set off (ignite), set up (arrange), set up (trick), set up (introduce), set right (rectify), set the stage (prepare)

put (cause to be at), put off (procrastinate), put off (offend), put one

over on (fool), put down (insult), put down (euthanize), put in for

(request), put out (extinguish), put out (inconvenience), put out

(consent to sex)

stand (rise), stand out (impress), stand up for (defend), stand in

(replace), stand off (repel)

But in every instance they retain their irregular past tense forms in the

extended meanings: Barney came around, Barney came out, Barney came

off as (never comed); Joan took him in, Joan took a bath, Joan took over

(never taked); and so on. All the meanings march in lockstep with the same

irregular past-tense forms, no matter how tenuous the semantic thread that

links them. The mind links an irregular sound such as took not with the

meaning of a word directly but with the word’s root—a unique address in

the mental lexicon, like the boldfaced entry for a word in a dictionary,

which can have several meanings listed under it.23

An even more curious demonstration comes from families of words

with the same stem and different prefixes. Words with prefixes keep the

past-tense form of the stem: eat—ate becomes overeat—overate; make—

made becomes remake-remade. That is not surprising, because we all hear

the eat inside overeat—overeating is, after all, a kind of eating, namely,

eating too much. What is surprising is that the same thing happens when the

meaning of the combination is opaque. Few people sense the meaning of

the stand inside understand, the get inside forget, or the come inside become. Nonetheless no one is tempted to say understanded, forgetted, or

becomed; the irregular forms persist, giving us understood, forgot, and

became. Here are some examples:

come—came, become—became, overcome—overcame

go—went, undergo—underwent

get—got, forget-forgot

take-took, mistake—mistook, overtake—overtook, partake—partook,

undertake-undertook

set—set, beset-beset, upset-upset

stand—stood, understand-understood, withstand—withstood

draw—drew, withdraw—withdrew

hold—held, behold—beheld, uphold—upheld, withhold—withheld

give—gave, forgive—forgave

Irregular forms stick like glue to their verb roots, even when reduced to

meaningless little tokens inside a bigger verb. Speakers of English seem to

analyze become as be- + come and understand as under- + stand, even though the meaning of become is not computable from the meaning of be-and the meaning of come, and understand has nothing to do with standing. This is not something we have to learn in school. When we acquire

language, our minds analyze sets of words, looking for their parts as if they

were clues in a combinatorial puzzle. We mentally arrange them in a matrix

according to overlap:

and use the common denominators in the rows and columns to make

incisions in the words, thinking of them thereafter as amalgams of parts:

become = be- + come, withdraw = with- + draw, and so on.24

Of course, it was English speakers of centuries past, our linguistic

ancestors, who first analyzed become as he- + come and extended the come

—came pattern to it, and it is possible that to them the words were as

transparently built out of parts as overeat or remake are to us today. Even

so, it is unlikely that we have been stupidly memorizing became, overcame,

withdrew, and so on as structureless strings of vowels and consonants. If we

were to come across a new complex word, such as undercome, bestand,

overhold, or withset, and were unaware of its meaning, we would almost

certainly use the irregular forms of the words inside them: undercame (not

undercomed), bestood, overheld, withset. Moreover, it is not pure sound that

carries the irregular form: The past of succumb and encumber are

succumbed and encumbered, not succame and encame-ber, because people

don’t perceive them as containing a prefix followed by the word come, only

the sound kŭm. Clearly the perception of an embedded word comes from its spelling:

become contains c-o-m-e; succumb doesn’t. But spelling does not directly inform speakers how to form the past tense; it merely assigns a distinct

visual signature to every root, and speakers choose the past-tense form that

goes with the root. Samuel Johnson, who standardized the spellings of

thousands of modern words, used people’s perception of the anatomy of

words as a rationale in his decisions, and that is one of the reasons that the

spellings of English words notoriously do not always reflect their sounds;

often they reflect morphological structure instead. We see this in the many

words that sound alike but are not perceived as being the same word (that

is, as having the same root), and are not given the same past-tense form:

In the last three cases the spellings divulge the presence of words that are

recognizable in other guises—the adjective bare, the noun steel, the noun

brake—and we will see in chapter 6 that this makes an especially big difference in how we compute their past tense forms.25

The English system of inflection, we have seen, dissects cleanly into a few

simple components. The past-tense rule belongs to a component,

morphology, that builds things out of parts using rules. The rule itself is a

masterpiece of minimalism—“a word can be composed of a stem and a

suffix”—with all other details distilled out and collected in the lexical entry

for the suffix. The suffix itself is shared among several inflections (past

tense, participle, and so on), and its variant pronunciations ( t, d, - d) do not

wastefully multiply listings but are computed automatically by two

ubiquitous rules of phonology. The distinction between the lexicon

(including irregular inflection) and grammar (including regular inflection)

is a distinction between a list of entries and an algorithm for combining

them, rather than a side effect of a general yearning to distinguish

meanings. That leaves the irregular words. Every irregular tells a story, and they

are the topic of the next chapter. 3

BROKEN TELEPHONE

In the game known as Broken Telephone (or Chinese Whispers) a child

whispers a phrase into the ear of a second child, who whispers it into the

ear of a third child, and so on. Distortions accumulate, and when the last

child announces the phrase, it is comically different from the original. The

game works because each child does not merely degrade the phrase, which

would culminate in a mumble, but reanalyzes it, making a best guess about

the words the preceding child had in mind. All languages change through the centuries. * We do not speak like

Shakespeare (1564-1616), who did not speak like Chaucer (1343–1400),

who did not speak like the author of Beowulf (around 750–800).

As the

changes take place, people feel the ground eroding under their feet and in

every era have predicted the imminent demise of the language. Yet the

twelve hundred years of changes since Beowulf have not left us grunting

like Tarzan, and that is because language change is a game of Broken

Telephone. A generation of speakers uses their lexicon and grammar to produce

sentences. The younger generation listens to the sentences and tries to infer

the lexicon and grammar, the remarkable feat we call language acquisition. The transmission of a lexicon and grammar in language acquisition is fairly

high in fidelity—you probably can communicate well with your parents and

your children—but it is never perfect. Words rise and fall in popularity as

the needs of daily life change, and also as the hip try to sound different from

the dweebs and graybeards. Speakers swallow or warp some sounds to save

effort, and enunciate or shift others to make themselves understood. Immigrants or conquerors with regional or foreign accents may swamp the

locals and change the pool of speech available to children. Children, for their part, do not mimic sentences like parrots but try to

make sense of them in terms of underlying words and rules. They may hear

a mumbled consonant as no consonant at all, or a drawn-out or

mispronounced vowel as a different vowel. They may fail to discern the

rationale for a rule and simply memorize its outputs as a list. Or they may

latch on to some habitual way of ordering words and hypothesize a new rule

to make sense of it. The language of their generation will have changed,

though it need not have deteriorated. Then the process is repeated with their

children. Each change may be small, but as changes accumulate over

centuries they reshape the language, just as erosion and sedimentation

imperceptibly sculpt the earth. That is how irregular forms, in particular, come down to us. Most of the

forms were originally created by rules, but a later generation never grasped

the rules and instead memorized the forms as words. They were words for

every generation thereafter, and each irregular was free to accumulate its

own quirks from subsequent distortions and reanalyses. Because irregulars

originated from rules they are not a random grab-bag but rather display

patterns, fossils of the long-dead rules. A. L. Kroeber, a founder of modern

anthropology, reminisced that his “first remembered purely intellectual

pleasure” was seeing patterns in English irregular verbs, a foretaste of his

search for systematicity in culture more generally. 1

This chapter is a guided tour of the irregular nouns and verbs of

English, with commentary on where they came from and where they are

going. These words all will have their turns on stage throughout the book,

so it’s helpful to get to know them individually. This is also a lively way to

come to understand how language changes, including how it is changing

today. People often ask me how linguists know the way people pronounced

things in centuries past. After all, Chaucer, unlike Nixon, did not secretly

tape his conversations for the benefit of future historians. Old

pronunciations can be painstakingly inferred from a diverse set of clues. One of them is spelling. Before Samuel Johnson standardized English

orthography, people spelled more or less as they pleased, trying to capture

the sounds of language as they heard them. Spellings were more phonetic,

and changes in spelling give clues to changes in pronunciation. For

example, when writers started to spell Old English bi-healfe (behalf) as

behaf, one can guess that people had stopped pronouncing the l. Other clues

come from wordplay. For example, Shakespeare rhymed or punned case

and ease, hate and eate, say and sea, and shape and sheep, suggesting that speakers of Early Modern English pronounced the vowels in each pair in

the same way (clues from spelling suggest it was ā).

A third kind of clue is

found in the writings of language snobs who criticize or lampoon the

speech of their contemporaries, inadvertently immortalizing it to the good

fortune of modern linguists. Other clues exist as well, and together they can

triangulate on the most common and most probable pronunciations. We can never say for sure what the pronunciation of a given word at a

given time actually was. Just as there are regional accents today (London,

Boston, Texas, and so on), there were regional varieties of English centuries

ago; indeed, many more of them, because people did not move around as

much as we do, did not send their children to melting-pot schools, and had

no dictionaries to consult. Also, the written record is haphazard. Most

words and pronunciations were in use long before the first literate person

chanced to write them down, and many others went to the grave along with

their speakers. When word histories can be reconstructed, invariably they

are convoluted, eye-glazing yarns. This is to warn you that the word

histories presented here have been simplified to highlight the kinds of

psychological processes that cause words to have histories. 2

Words aren’t regular or irregular across the board. Words are regular or

irregular only with respect to certain inflections, some more tolerant of

irregularity than others. The present progressive suffix -ing, as in The joint is jumping, is 100

percent regular. There isn’t a single exception to the rule, not even the

rebellious be, which meekly submits and shows up as being. Why, when it

comes to -ing, does no verb hear a different drummer?

One reason is that

the progressive construction came into English relatively recently, late in

the Middle English period of 1100 to 1450.

It borrowed the -ing suffix from

the gerund (a construction that turns a verb into a noun, as in the changing

of the guard), and the newly cloned -ing suffix had the progressive all to

itself and did not have to compete with alternative forms hanging around

from earlier periods. Another reason is that -ing is found in a separate

syllable, which makes it easy for listeners to hear a word such as breaking

as break + ing. That is an advantage over -s and -ed, which can sound as if

they are part of a stem, like act, box, or maze. As we shall see, the

camouflage of -s and -ed can invite listeners to mis-analyze a regularly

inflected combination as a one-piece irregular word. One other suffix is completely regular: the possessive s. Any noun can

take it, even the irregular nouns that cannot appear with an s sound when it

is a plural suffix, such as mouse and man. We have no trouble saying the

man’s hat, the mouse’s mother, or the goose’s egg, even though we never

say the mans, the mouses, or the gooses. Why no irregulars?

The possessive

is unusual because it attaches to a phrase rather than to a word. One can talk

not just about the cat’s pajamas but about the cat in the hat’s pajamas,

where the pajamas belong to the cat, not to the hat:

A former student, Annie Senghas, once said to someone at a conference,

“The woman sitting next to Steven Pinker’s pants are like mine.” I was fully

clothed; the woman sitting next to me had pants like Annie’s. Dave Barry’s

column-within-a-column “Ask Mr. Language Person once had the

following exchange:

Q: Recently, did your research assistant Judi Smith make a

grammatically interesting statement regarding where her friend,

Vickie, parks at the Miami Herald?

A: Yes. She said, quote, “She comes and parks in whoever’s not

here’s space that day.”

The word here is not even a noun!

Since ’s is perceived not to be attached to

an adjacent noun, it cannot unite with that noun in people’s minds, and

therefore never evolves into an irregular word. The exceptions that prove

the rule are the possessive pronouns my, your, his, her, our, and their, which

are, in a sense, irregular replacements for me’ s, you’s, him’s, her’s, us’ s, and them’s. Pronouns are one-word phrases; in any sentence position where you

can say the man in the gray suit you can also say he or him. A pronoun, being a phrase, is the only kind of word that could form a cohesive

amalgam with ’ s, which in effect is what possessive pronouns are. The third-person singular -s, as in Dog bites man, steps aside for

irregular forms in only four verbs: be—is (not be’s), have—has, do—does

(pronounced dŭz), and say—says (pronounced sĕz).

These, by the way, are

the four most frequent verbs in the English language. 3 In chapter 5 we will see that this is not a coincidence. Nouns embrace several kinds of irregular plurals. 4 Many nouns

ordinarily don’t take any plurals: mass nouns such as mud, celery, furniture,

and evidence are treated as seamless stuff rather than countable things. (A

former graduate student who is a Russian emigré was teased by fellow

students for saying, “I hev three evidences for thees theory.”) Of the count

nouns, which do take plurals, exactly seven change their vowel instead of

adding - s:

man-men, woman—women (pronounced wĭm n) , foot—feet, goose

—geese, tooth—teeth, mouse—mice, louse—lice

Why do we flip the vowels in these nouns?

Originally they took plural

suffixes, just like regular nouns, though the suffixes were different from

today’s -s. For example, foot, originally fot, had the plural foti. But as we

saw in chapter 2, you can’t just force a consonant or vowel onto the end of a

word and hope that nothing else happens. People adjust their pronunciation

of a sound in anticipation of the sounds to come. In many modern English

dialects, for instance, speakers pronounce the i differently in write and ride

and the ou differently in shroud and about. In keep cool the first k sound is pronounced toward the front of the mouth, the second one toward the back. In words like find and sound the n vanishes and the vowel reminds us of the

vanished consonant by being sounded through the nose. Most of us are

unaware that we make these adjustments and are puzzled when children

spell find as fid, though it is an accurate transcription of the n-less word they hear. Some of these adjustments come from the way we control our

muscles, but others get standardized into phonological rules, which define

what we hear as an accent. In the Germanic languages that were ancestral to English there was a

phonological rule that changed the pronunciation of a vowel from the back

of the mouth to the front of the mouth if the next syllable contained a vowel

pronounced high and in front. The rule spared people from having to jerk

their tongue backward and then forward while pronouncing the words. So

in foti, the plural of fot, the back o was altered to a front e, harmonizing with the front i in the suffix: roughly, feti. The process is called umlaut and

it is still visible in our linguistic cousin, German, as the two little dots over

some vowels: die Kuh (the cow), die Kühe (the cows).

In the Middle English period, speakers began to mumble the unstressed

syllables at the ends of words and then began to drop them outright. At that

point people must have been hearing the altered vowel in feti as a different

vowel altogether, not as a tweaked o, because when the suffix was dropped,

the speakers kept the altered vowel in the stem, even though nothing was

there to tweak it anymore. The eventual result was feet. It reminds me of the

explanation of why there is a basketball team in arid Los Angeles called the

Lakers and a team in pious Utah called the Jazz. Originally the teams were

based in Minneapolis, The Land of Lakes, and in New Orleans, The

Birthplace of Jazz. When the teams moved, they kept their names, even

though the names no longer made sense. Another three irregular plurals take the old Anglo-Saxon suffix -en

rather than -s:

child—children, ox—oxen, brother—brethren

Of the three, only children is part of the standard American vernacular

(though the others are preserved in some nonstandard dialects, together with

archaic plurals such as eyen, shoon, and hosen).

Most Americans meet oxen

mainly in writing, and commonly say oxes instead. 5 Similarly, they perceive brethren as an inkhorn term for monks and parishioners. As a

result, the -en sounds archaic and lends itself to silly wordplay. Shortly after

the appointment of Ruth Bader Ginsburg to the U. S. Supreme Court, where

she joined fellow person of gender Sandra Day O’Connor, Newsweek

reported, “The brethren—and now, two sistren—had to know that the

swiftness and scope of their ruling would be viewed as a landmark victory

for working women.”6 In the argot of computer hackers, who try to outdo

each other with logical extensions of irregular patterns, the plural of the

computer called the VAX is VAXen, and there also have been sightings of

faxen, boxen, soxen, and Macintoshen.7

Several names for gregarious animals that are hunted, gathered, or

farmed are identical in the singular and plural:

fish, cod, flounder, herring, salmon, shrimp

deer, sheep, swine, antelope, bison, elk, moose

grouse, quail

These forms may have come from constructions in which the singular is

used to refer to potential quarry in the aggregate, as in We went hunting for

duck. A fourth class of nouns takes the regular -s ending but changes its final

consonant, usually f but sometimes th or s, from unvoiced to voiced:

calf—calves; also elf, dwarf, half, hoof, knife, leaf, life, loaf, self,

scarf, sheaf, shelf, thief, wife, wharf, wolf

mouth—mouths; also truth, sheath, wreath, youth

house—houses

Something familiar is going on here: A voiced consonant z is being shoved

against an unvoiced consonant, and one of them bends to make the cluster

consistent. We saw this happening in the regular nouns, where -s is

pronounced differently in dogs and cats. But strangely, in these nouns the

suffix z keeps its voicing, and the noun surrenders it—a right-to-left

smearing that violates the usual left-to-right smearing of English

phonology. Some linguists have posited a special rule, regressive voicing, to

generate these examples. The rule, though, would have to be handcuffed to

these two-dozen-odd words, because most nouns ending in f or th are

regular and would have to be left untouched. The plural of reef is regular

( reefs, not reeves), and the same is true for nouns such as these:

birth, booth, earth, faith, growth, hearth, length, month, tenth

belief, brief, chief, proof, safe, spoof, turf

Even many of the so-called irregular nouns are questionable; many speakers

simply pronounce hoofs, wharfs, oaths, and truths in the ordinary way. I

prefer a different theory: that some nouns have two stems, one for the

singular, one for the plural, and that the plural stem is tagged as incomplete

without a suffix: knive-, loave-, wolve-, and so on. After all, if -ed and -ing

are tagged as suffixes that cannot be pronounced unless they are attached to

a stem, why can’t there be stems that cannot be pronounced unless they

have a suffix attached to them?

The regular suffix -s then applies, generating the plural form without further ado. 8

Finally, there are nouns that take Latin or Greek plurals. As the singer

Alan Sherman has pointed out, “One hippopotami / Cannot get on a bus. Because one hippopotami / Is two hippopotamus.” Here are five families

with Latin plurals:

alumnus—alumni; also bacillus, cactus, focus, fungus, locus,

nucleus, radius, stimulus

genus—genera, corpus—corpora

alga—algae; also alumna, antenna, formula, larva, nebula, vertebra

addendum—addenda; also bacterium, curriculum, datum,

desideratum, erratum, maximum, medium, memorandum,

millennium, moratorium, ovum, referendum, spectrum, stratum,

symposium

appendix—appendices; also index, matrix, vortex

And here are two families with Greek plurals:

analysis—analyses; also axis, diagnosis, ellipsis, hypothesis,

parenthesis, synopsis, synthesis, thesis

criterion—criteria; also automaton, ganglion, phenomenon

These nouns come from science and academia, and the plurals were

borrowed directly from Latin or Greek together with the singulars. They

must be irregular forms that are memorized as a list, not the products of a

rule attaching -i or -ae, because most nouns shun these plurals except in the

speech of people with an attitude:

apparatus—apparatuses; also bonus, campus, caucus, census,

chorus, circus, impetus, prospectus, sinus, status, virus

area—areas; also arena, dilemma, diploma, drama, era, etc. album—albums; also aquarium, chrysanthemum, forum, museum,

premium, stadium, udtimatum

Latin- and Greek-inspired plurals in a sense are still not part of the

English language. They are not acquired as part of the mother tongue in

childhood, and are uncommon in everyday speech among nonacademic

adults. Instead they are learned in school together with the Pythagorean

theorem and the dates of the Peloponnesian War. Since they follow no

living rule, and people couldn’t have memorized them unless they went to

the right schools and read the right books, they are shibboleths of

membership in the educated elite and gotcha!

material for pedants and

know-it-alls (the kind of people who insist that the millennium begins

January 1, 2001).

Admittedly, I cringe when I hear this phenomena, those criterias, and

the media is, and I could barely contain myself during the speech from the

president of the alumni association who kept thanking the alumnis. I also

get a perverse pleasure from correcting students who refer to an important

piece of data or write that this data is important. ( Data is the plural of datum, I tell them, so one ought to say, The datum is important; The data

are important. ) Yet by the same logic I ought to correct myself when I refer

to an agenda, two candelahras, this insignia, or that propaganda, which are

the plurals of agendum, candelabrum, insignium, and propagandum. And I

refuse to hear a word about genii, termini, aquaria, podia, lexica, fora,

stadia, or apices. In any case, whenever pedants correct, ordinary speakers

hypercorrect, so the attempt to foist “proper” Greek and Latin plurals has

bred pseudo-erudite horrors such as axia (more than one axiom), peni,

rhinoceri, and this one:

“Fellow octopi, or octopuses...

octopi?...

Dang, it’s hard to start a

speech with this crowd.”

THE FAR SIDE© 1987 FARWORKS, INC. Used by permission. All rights

reserved. It should be “Fellow octopuses.” The -us in octopus is not the Latin noun

ending that switches to -i in the plural, but the Greek pous (foot).

The

etymologically defensible octopodes is not an improvement. The flip side of plural pomposity is playful punning that deflates it, and

for decades wags have seen the opening. In a Peanuts cartoon, Linus had to

bring eggshells to Miss Othmar’s class so he could make igli. The comedian

Shelley Berman has talked of stewardi wearing blice. Wayne and Shuster

performed a skit in which Julius Caesar nibbled on a spaghettus. In Richard

Lederer’s “Foxen in the Henhice,” Farmer Pluribus reached for some

Kleenices while being serenaded by tubae, harmonicae, accordia, fives, and

dra.9 Henry Beard and Roy McKie’s A Gardener’s Dictionary contains the

following entry:10

Narcissus: wonderful, early-blooming flower with an unsatisfactory

plural form. Botanists have been searching for a suitable ending for

years, but their attempts —narcissi (1947), narcissusses (1954),

narcissus for both singular and plural (1958), and multinarcissus and

polynarcissus (1962, 1963)—haven’t enjoyed any real acceptance,

and thus, gardeners still prefer to plant the easily pluralized daffodil

or jonquil. This may seem silly and inconsequential, but the following story

appeared in The New Republic on December 12, 1994: “In Las Vegas, The

Flying Elvi sued The Flying Elvises for trademark theft. Both organizations

leap from airplanes in Elvis Presley (late period) costumes and dance and

pretend to sing upon landing.”

The masterpiece in the underappreciated genre of irregular plural humor

comes from the National Puzzlers’ League, the association of twisted

geniuses who devise impossibly clever word problems. One kind of puzzle,

the falsie, begins by finding a pair of words that look as if they are related

by a morphological rule:

False iteration: bus—rebus, bozo—rebozo, ally—really

False predecessor: lope—antelope

False feminine: butter—buttress, car—caress, under-undress

False comparative: ling—linger

False plural (from Hebrew): inter—interim

The puzzle itself takes the form of a poem (called a flat) that uses a pair of

falsely related words. The words are deleted from the poem and their

locations are marked with placeholders. The object of the puzzle is to guess

the pair of words from the context of the poem. The following flat by the

puzzler known as Trazom (in real life Joshua Kosman, the senior music

critic of the San Francisco Chronicle) contains a seven-letter singular noun

in the place marked one and its six-letter false plural in the place marked

many. Try it (the answer is in the notes).

False Plural (7, 6)

Turn over on your side, my dear,

And tuck your foot behind your ear;

And I, meanwhile, will crouch like this

And give your neck a tender kiss. Let’s see now—let your arms go slack

And clasp your hands behind my back;

I’ll reach around and drape my knee

Across your shoulder—goodness me!

I must confess, this is a stretch,

But honeybunch, you mustn’t kvetch. I know it hurts, I know it smarts—

But these arcane erotic arts

Don’t yield their secrets right at first;

And now, I think, we’re past the worst. So please don’t throw a ONE, sweet miss—

The MANY says we’ll soon reach bliss.11

Now we come to the irregular verbs. A menagerie of nearly two hundred

words coming in many shapes and sizes, they are a vivid demonstration of

how the human mind, reacting to the events of history, reshapes a language

over centuries and millennia. 12

The verbs he, have, do, and go are irregular in many of the world’s

languages. They are the most commonly used verbs in most languages and

often pitch in as auxiliaries: “helper” verbs that are drained of their own

meanings so that they may combine with other verbs to express tense and

other grammatical information, as in He is jogging, He has jogged, He

didn’t jog, He is going to jog. Many language scientists believe that the

meanings of these verbs—existence, possession, action, motion—are at the

core of the meanings of all verbs, if only metaphorically. For example, the

mind treats telling him a story as causing the story to go to him resulting in

him having it, and it treats dying as going out of existence.13

In English we saw that be stands out from all other verbs with its eight-

way conjugation. Its irregular past-tense form stands out too. Together with

go, it is the only verb whose past tense is a completely unrelated word, a

relation that linguists call suppletion:

be—was/were—been

go—went—gone; also undergo, forgo

Suppletion arises from a merger of two verbs. Old English, spoken from

about 400 to 1100, had three verbs for be: beon, esan, and wesan. They

probably differed in meaning, with beon referring to permanent states and

the other be s to temporary ones. (The distinction is similar to the one in

modern Spanish between ser and estar: Yo soy Americano \[I am American],

a long-term trait, contrasts with Yo estoy contento \[I am happy], a temporary

state.) Adding to the surfeit, different sets of be s were used in different parts

of England. In the Middle English period (1100-1450) they merged into one

verb. As in a corporate merger, in a linguistic merger the workers scramble

to fill a smaller number of positions, because a verb generally permits only

one form in every slot in its conjugation. Beon supplied the base form be;

esan supplied am, is, and are; wesan supplied was and were. For mysterious reasons, in the Middle English period the verb go

usurped the past-tense form of another verb, wend (as in to wend one’s

way), namely, went. Today the verb wend, bereft of its old past-tense form,

has the regular past wended, but its original form followed a pattern that can

be seen today in other irregular verbs, such as bend—bent, send—sent, and

spend—spent. Have, also irregular in many languages, is one of two English verbs that

drops its final consonant and replaces it with a d:

have—had, make—made

Originally these were haved and maked, but enough lazy speakers

swallowed the consonants that at some point in the Middle English period

speakers didn’t hear them and assumed that they were not there at all. The verb do does something slightly different—it takes on a -d, and

changes its vowel: do—did—done. Its participle form done (as in You’ve

done it again) is a contraction of the verb with an old suffix, -en. (The same

thing happens in be—was—been and go—went—gone. ) The -en suffix is

found in about fifteen English participles (such as spoken, sworn, chosen,

blown, and written), but the suffix is not attached by a rule. New verbs, such

as the neologisms fax, Bork, spam, and mosh, never get -en participles; no

one says:

I’ve already faxen it. That’s the third nominee the Republicans have Borken this session. The company has spammen its customers with ads once too often. Not tonight, dear; I’m sore from having moshen all night. Putting aside weird be, what do all these verbs— had, made, did, and the

bent—sent—spent family—have in common?

They all end in t or d. These,

of course, are the same consonants that make up the pronunciation of the

regular suffix -ed. About half the irregulars end in t or d, because they originally took some version of the regular -ed suffix but then fell off the

regular bandwagon for one reason or another. These lapsed regulars,

together with the regulars themselves, were dubbed weak, in 1819 by Jacob

Grimm of Grimms’ Fairy Tales fame; Grimm was also one of the first

historians of the Germanic languages. Grimm called the verbs “weak”

because they were too wimpy to hold on to their own unique past-tense

forms. We will meet the more macho strong verbs later in the chapter. Some version of the weak past-tense suffix -ed can be found in all the

Germanic languages, including English, German, Dutch, and the

Scandinavian languages. The suffix originated in an ur-language, Proto-

Germanic, spoken by a tribe that occupied most of northern Europe in the

first millennium B. C. Linguists call it the dental suffix because it was

pronounced with the tongue against the gum ridge behind the teeth. Why didn’t the weak verbs make life simple and just stay regular?

It is

because combinatorial rules of grammar have a cost, as we saw in chapter

1: They blindly join things together without looking at what they are made

of, and thus can create ungainly chimeras. Two strange things can happen

when a verb finds itself with a suffix grafted onto its rear end. One of them

is illustrated by the largest class of irregular verbs in English, the no-change

verbs:

hit—hit; also slit, split, quit, knit, fit, spit, shit

rid, bid, forbid

shed, spread, wed

let, bet, set, beset, upset, wet

cut, shut, put

burst, cast, cost, thrust

hurt

(Some of these verbs have alternative past-tense forms: irregular bid—bade,

forbid—forbade/forbad, spit—spat, the mainly British shit—shat, and

regular slitted, knitted, fitted, wetted, and thrusted.)

Note that all twenty-eight verbs end in t or d. Most of them arose in

Middle English and Early Modern English (1450-1700) when the regular

ending was often -de or -te. Throughout the language, es at the ends of words, formerly pronounced, were dropping like flies; the “silent e” in the

modern spelling of words such as bake is a souvenir of the earlier period. Thus a form such as hitte got reduced to hit. But why did speakers stand by

as these past-tense forms shrank into confusing copies of their stems, rather

than making the verbs regular, which would have given them the more

distinctive hitted?

If I may be permitted to psychoanalyze speakers who have been dead

for centuries, it probably came from a widespread human habit: We don’t

like to put or keep a suffix on a word that looks like it already has the

suffix.14 In this case, people don’t like to put a version of -ed on a verb that already ends in t or d. Psycholinguists have offered several explanations. Perhaps speakers develop a stereotype for “past-tense form, namely, “ends

with t or d, ” and unconsciously think that a stem that fits the stereotype has

already been inflected and stop themselves from adding the suffix again. Perhaps when the mind assembles past-tense forms, it gets confused

between the it or ed or ut that is already at the end of the stem and the -i or -

d it is trying to add and merges them into a single sound, like the girl who

said, “I know how to spell banana, but I don’t know when to stop.” Perhaps

the suffix -d is applied, and the unpronounceable result, hitd, is cleaned up,

not by the ordinary phonological rule that inserts a neutral vowel between

the t and d, but by a special rule that deletes the d. Perhaps several of these explanations are correct. In any case the no-extra-suffix habit is alive and well in modern

speakers. The psycholinguists who jot down speech errors have found that

people are prone to leaving out -ed on regular verbs that end in t or d. For example, they say, So we test ’em on it, intending to say tested, or That’s what I need to do, intending to say needed. 15 The same thing happens when people are brought into the lab, given a list of verbs, and asked to say them

aloud in the past tense as quickly as they can. 16 Children, too, don’t like to

add -ed to verbs ending in t or d—they make their signature error, breaked,

less often with verbs that end in t or d, such as hitted, putted, builded, and meeted, than with verbs with other endings, such as bringed and buyed. 17

These habits are leaving their mark on English as it continues to evolve:

Even in careful speech and writing, many people use no-change past and

participle forms for verbs like bust, pet, shred, and tread, as in She got the

fleas when she pet the dog and This is an area where few psychologists have

tread. 18

The phobia of adding a surplus suffix extends beyond the past tense. Gardening scriveners often cannot bring themselves to write crocuses,

gladioluses, and narcissuses (as we learned in the Gardeners’ Dictionary

entry for Narcissus), and write headlines such as “Hardy Gladiolus Have

Long Been a Favorite,” and “Dutch Crocus Herald the Arrival of Spring.”

(No doubt these are symptoms of a Latin-conscious -us/ -i anxiety as well.) I

have seen an ad for a sprayer that fits all hose and another one for the

pantyhose that last, and still another announcing All fax on sale. People

treat the sh sound as similar to s, leading the Boston Globe handyman to

write about adjusting window sash,19 and leaving every professor baffled as to how to refer to more than one prefrosh (pre-freshmen).

Many people

have trouble keeping up with the Joneses and instead merely try to keep up

with the Jones. When it comes to the possessive s, hardly anyone follows

the advice in Strunk and White’s famous style manual to refer to Charles’s

hat (charlz z) or the Jones’s car; it’s usually Charles’ hat and the Jones’

car, both in writing and speech. And what do you say to someone who has a

daddy-long-legs climbing up each shoulder?

When a word has a verbatim replica of a suffix inside it, rather than just

a reminder of one, the attempt to add the real suffix often results in

clumsiness or unintelligibility. When there is rain or snow or hail or thunder

coming down from the skies, it is said to be raining, snowing, hailing, and

thundering. What about when there is lightning?

Is it lightninging?

Not very

likely, and some speakers snip out an -ing and say It is thundering and

lightning. Many adjectives can be turned into adverbs by adding -ly, such as

softly, surely, and happily. What about those adjectives that already end in -

ly, such as ugly, friendly, heavenly, or leisurely?

Uglily?

Friendlily?

Heavenlily?

Leisurelily?

Pthack. (The Atlantic Monthly, perhaps hoping no

one would notice, once ran a story entitled “Friendily Yours.”) Sometimes

brand names can be turned into colloquial verbs for traveling or sending:

We Chevy’d up and down Main Street. I FedExed the package last night. Down to their last thirty dollars, they Greyhounded home. Because of his fear of flying he Amtrak’d to New York. But even if your frequent flyer plan is with United Airlines, it is unlikely

that you have ever Uniteded to San Francisco. Sometimes people can get into trouble by speaking as if a word that

appears to contain an affix really does contain it. An interstate trucking

company must have lost the business of the literate when it proudly painted

its trucks with the slogan “Faster than rail, regular than mail. ”20 Former President George Bush used to tell reporters that he spent his vacation

bonefishing, leading them to wonder what the best bait is for catching

bones, and presumably his heart was in the right place when he explained,

“I hope I stand for anti-bigotry, anti-Semitism, anti-racism.”

Back to the verbs. Repeated-suffix phobia is also the explanation for the

class that originally contained wend—went:

bend-bent; also send, spend, lend, rend, build

These verbs devoice their final consonant, d, into t. They began as bend +

de, and the double d was fixed by trimming the final consonant of the stem,

yielding past tense ben + de. The extra twist is that the phonological rule

that today turns -d into -t in words like walked and passed used to be triggered by words ending in l, m, n, and v as well. Bende became bente, which then lost its e to give us bent. The overeager -d → -t rule can also be blamed for these verbs:

burn—burnt; also learn, dwell, spell, smell, spill, spoil

The irregular forms ending in -t show the English language changing before

our eyes: Most of them are on their way out. American speakers mainly use

burnt as an adjective, not a past-tense form— The toast is burnt because

Bernie burned it—and would not be caught dead saying learnt, dwelt, spelt,

smelt, spilt, or spoilt. Rent is used only for emotional resonance, as in The

Vietnam war rent the fabric of American society, and lent is giving way in

American English to loaned, the past of to loan. In “Childe Harold’s

Pilgrimage” (1812) Byron describes a battlefield using three verbs in the

class that range from the moribund to the dead:

The thunder clouds close o’er it, which when rent,

The earth is covered thick with other clay

Which her own clay shall cover, heaped and pent,

Rider and horse—friend, foe, in one red burial blent!

Blend—blent, of course, has become completely regular, as have most of

the other verbs with -t in their past-tense forms, such as wend—went, pen—

pent, gird—girt, geld—gelt, and gild—gilt. Like many obsolete irregulars,

gilt and pent have left relics among the adjectives: a gilt-edged book, pent-

up energy. Another reason that regular forms can go to seed is the Los Angeles Lakers

effect that gave us irregular plurals such as feet and mice. Grafting a suffix

onto a stem can trigger changes in the pronunciation of the stem, and

sometimes the change can stay in the word long after the trigger has

vanished. Many languages distinguish a vowel sound pronounced quickly from

the same sound drawn out; they are called short and long vowels. The

vowels traditionally called “short” and “long” in English, such as the ones

in bet and beet, used to differ in this way, as we see in their spellings: The

long vowel was symbolized by writing two short vowels in a row, as if it

took twice as long to pronounce. Starting around the year 1000, English speakers shortened their

pronunciation of a vowel when extra phonetic stuff (such as a consonant or

syllable) was added, pushing new consonants into the syllable. 21 Here are

some examples that have survived in modern English:

bone—bonfire

break—breakfast

child—children

Christ—Christmas

deep—depth

five—fifth

know—knowledge

sheep—shepherd

wide—width

wise—wisdom

Shortening a vowel is a natural reaction when material is added to the end

of the syllable. A syllable is a unit of timing, taking up a constant tick of the

speech clock. If material is added to the end of a syllable, the vowel is often

shortened to maintain the rhythm.22 This habit of pronunciation could

easily have turned into a full-fledged rule. In his April 5, 1997 column, the

language maven William Safire ventured that the pronunciation of seminal

as “SEM-uh-null” in place of “SEE-muh-null” was an instance of academic

bowdlerization—prissy professors covering up the fact that the word

seminal comes from the word semen. Safire’s theory, however, would have

to go to lengths worthy of Oliver Stone to explain why those professors,

presumably hatching plots in their SEEminars, have also changed the

pronunciations of vanity, sanity, cleanliness, brevity, and criminal to hide

the fact that they come from vain, sane, clean, brief, and crime. All, of course, are products of a phonological rule in English that shortens vowels

at the beginning of many three-syllable words. Take a verb with a long vowel like keep. Add the regular suffix and

spell it phonetically: keept. Shorten the vowel in response to the extra stuff

at the end. We end up with something pronounced kept—one of a number

of modern irregular past-tense forms that would be regular but for their

shortened vowels:

keep—kept; also creep, leap, sleep, sweep, weep

Add some other habits of Middle English speakers that we have come

across—using -t more widely, dropping suffixes—and you understand

many other irregular verbs in modern English:

feel—felt; also deal, kneel, dream, leave

bleed—bled; also breed, feed, lead, mislead, plead, read, speed,

meet

hide—hid; also slide, bite, light, alight

flee—fled, say—said, hear—heard, lose—lost, shoot—shot

sell—sold; also tell, foretell

do—did

(As before, some of these verbs allow regular past-tense forms, such as

kneeled, dreamed, speeded, lighted, and especially, pleaded. For some of

the verbs—particularly sell, tell, and do—the reasons for the vowel changes

are a bit more complicated.)

Kept, of course, isn’t simply keeped pronounced with a clipped vowel;

neither is hid just a short version of hide nor shot a short version of shoot. The pairs of vowels traditionally called “long” and “short,” and spelled as if

they are double and single scoops of the same sound, are in fact very

different vowels. How did that happen?

The perpetrator is a process of language change that is the opposite of

the various slurrings and swallowings and cutting of corners that we have

seen so far. All of those changes make it easier for the speaker to speak but

do nothing for the listener, who would rather have the speaker enunciate

clearly. Sometimes listeners do get their way; speakers enhance the

difference between a pair of vowels by adding, exaggerating, or

embroidering each in a different manner. 23

For many centuries speakers of Old and Middle English enhanced the

difference between short and long vowels by making the long vowels tense:

that is, the muscle at the root of the tongue is tensed up, changing its shape

and making the vowel in great sound different, as well as longer, than the

vowel in get. Enhancement went wild, however, during the dawn of Early

Modern English in the fifteenth century, when the pronunciation of the long

vowels was scrambled in a linguistic revolution called the Great Vowel

Shift. Before the shift, keep had been pronounced something like cape, hide

like heed, boot like boat. After the shift, the English spelling of the long vowels no longer made much sense, nor did the pairings of “short” and

“long” vowels in siblings like keep and kept. Since the children of Early Modern English could not have heard a relationship between the vowels,

the past-tense forms struck them as a ragbag that just had to be memorized

outright, and so they remained for subsequent generations. Thus verbs that

entered the popular language after the Great Vowel Shift, such as peep

(1460) and seep (1790), and verbs whose pronunciations eventually drifted

into rhyming with the keep verbs, such as reap and heap, did not undergo a

vowel change; they remained intact when they first submitted to -ed, giving

us peeped, seeped, reaped, and heaped, not pept, sept, reapt, and heapt. Here is a small mystery: What is the verb that goes with the past-tense

form wrought, as in The Watergate scandal wrought great changes in

American politics, and the participle form in Judges 23:23, Whàt hath God

wrought!

, quoted by Samuel Morse in the first intercity telegram?

According to the theory that irregulars are pairs of memorized words, an

irregular past-tense form could, in principle, survive in memory without a

corresponding stem. Wrought appears to be an example: Most people have

no idea what the verb is. Many guess wreak (based on an analogy with seek

—sought) or wring (based on an analogy with bring—brought), but both guesses are wrong. The answer is work: Wrought iron is worked iron, and a

person who is all wrought up is a person who is all worked up. (Old theater

saying: “Plays are wrought, not written.”) Wrought belongs to a family of

verbs that replace their rhyming parts with ought or aught:

buy—bought; also beseech, bring, catch, fight, seek, teach, think

How do you get wrought from work or sought from seek?

The

connection is less mysterious when we realize that the now silent gh used to

be pronounced, somewhat like the ch of Bach, loch, and Chanukah. Start with work (actually wryrcan, but I will use modern spellings to make the

changes clearer).

Add the suffix -t to get workt. Soften the k sound to gh, yielding worght—an old phonological trick to avoid the strenuous -kt. A

vowel and an adjacent r often switched places in the history of English,

because r sounds a lot like a vowel, which makes its order with respect to a

vowel hard to hear. Thus brid became bird, thrid became third, hross

became horse, and worght became wroght. We no longer pronounce the gh, and recall that many English vowels were shuffled during the Great Vowel

Shift (the vowel spelled ou was once pronounced ō), and that vowels often

get shortened when a suffix is added (so ō becomes ŏ).

The result is

wrought and the mystery is solved. In the 1980s the irascible New York Times book reviewer Anatole Broyard

wrote that he doubted that English had “any life left in it, any flavor or

idiosyncrasy.” His colleague Maggie Sullivan followed up in a column of

her own:

Anatole Broyard is right to sound the alarm. We are losing this

idiosyncrasy; as a language changes, strong verbs tend to become

weak. For example: although once shepherds shore their sheep,

sheep are no longer shorn, they are sheared. This issue should arouse lovers of the English language. Weakening

the verbs can only weaken the language itself. To keep English from

becoming a feeble tongue, we must reinforce our verbs. Fortunately,

I have come up with a two-part plan. First, we must not allow new

verbs to enter the language in a weak state. We must ensure, for

example, that to clone is established as clone, clewn, clown, as in:

Future generations of booksellers may reproach us for not having

clown Joyce Carol Oates and Isaac Asimov....

And to gentrify as

gentrify, gentrifo, gentrifum, as in: The newcomers gentrifo one

block and now the whole old neighborhood is gentrifum. Since new verbs are few and far between, I offer the second part

of my plan—creating new strong verbs. English has some strong

verbs with unique patterns for their principal parts, such as go, went,

gone. Individuality makes them particularly vulnerable. Their

patterns would hold up better if each pattern had more

representatives. If we create allies for our unique strong verbs, we

can buttress them and increase their number. Here are suggestions

for new strong verbs:

Conceal, console, consolen: After the murder, Jake console the

weapon. Subdue, subdid, subdone: Nothing could have subdone him the way

her violet eyes subdid him. Fit, fat, fat: The vest fat Joe, whereas the jacket would have fat a

thinner man. Displease, displose, displosen: By the look on her face, I could tell

she was displosen. Sullivan’s plan to “strengthen” the language captures two hallmarks of the

second kind of irregular verb in English, the so-called strong verbs. They

belong to alliances with similar sounds, and despite this solidarity, they

have been dwindling for millennia. The families of strong verbs have a history stretching back more than

5500 years. Most of the languages of Europe, Iran, and the northern half of

India, and many current and extinct languages of Turkey, western Asia, and

China, show similarities in vocabulary and grammar that suggest they are

descendants of a single language spoken by an expansive and mysterious

prehistoric tribe. The most popular theory is that they were a late-neolithic

farming people with domesticated horses, wheeled vehicles, and a military

leadership, who expanded from a homeland in southern Russia around 3500

B. C.24 An alternative is that they were the people that first brought farming

to Europe, beginning in 7000 B. C. from a homeland in eastern Turkey. 25

Though we don’t know who they were or where they came from, we know

a lot about how they spoke. Their language, Proto-Indo-European, has

largely been reconstructed by historical linguists working backward from

the commonalities in the daughter languages. 26

Many Indo-European languages have echoes of the strong-verb patterns

seen in English, such as bear—bore, tear—tore, and sink—sank, drink—

drank. Some of these verbs and their past tense forms actually existed in the

ancestral language, such as bher— bhor- and senkw— sonkw-.

Proto-Indo-European apparently had a set of rules for forming the past tense, not by

adding a suffix as in modern English, but by changing the vowels, as in

modern Hebrew—a kind of rule called gradation, apophony, or ablaut. There were probably seven ablaut rules, more or less as follows: If the verb

has ei followed by a consonant, change the ei to a. If the verb has e

followed by a vowel-like consonant, change the e to ă—and so on for the other five classes. When the Indo-Europeans started to spill out over Eurasia, the daughter

tribelets lost touch, and games of Broken Telephone began in each one. Eventually the language radiated into the ancestors of our familiar

languages and language families such as Germanic, Romance, Slavic,

Celtic, Greek, Iranian, and Sanskrit. For example, the verb werg- (to do)

ended up in Germanic as werkam (work), and in Greek as erg- (action) and

org- (tool), which eventually crossed over into English as energy, organ,

and orgy. When a word meaning “do” turns into a word meaning “orgy,”

the changes wrought by the chain of whisperers must have been

considerable. It is remarkable that the seven classes of Indo-European

strong verbs came through, tattered but recognizable, in Proto-Germanic,

then in the West Germanic language spoken by the Angles and Saxons, and

then in Old English, Middle English, and Modern English. That is why the

strong verbs fall into clusters of similar-sounding forms today. The rules themselves, however, did not survive. Imagine a rule that

replaced ĭ with a, and suppose that people started pronouncing ĭ as ī in some verbs, e in others, and ĭ in still others, depending on the consonants

following the vowel and many other factors. Children would have a hard

time making sense of the rule, and at some point they would stop trying and

simply memorize the past-tense forms as a list. By the time of Old English,

the Indo-European vowel-change rules were extinct and their products had

been mangled in different ways by the slings and arrows of outrageous

fortune. At least a fifth of the verbs no longer obeyed the rules of their

original class, and in the following centuries so many verbs joined, left, or

switched classes that today the classes no longer correspond very well to

the organization of the verbs in the minds of modern speakers. Here is one Old English class, Class I, that has come through in

recognizable shape:

rise—rose—risen; also arise, write, smite, ride, stride, dive, drive,

shine, strive, thrive

The list highlights a key feature of the strong verbs. While dictionaries

happily list irregular forms such as smite-smote—smitten, stride—strode—

stridden, strive—strove—striven, and heaven help us, thrive—throve—

thriven, in the minds of real English speakers these forms are muzzy:

People vaguely recognize them from books but are uncomfortable using

them in their own speech and are tempted to default to regular forms like

smited, strided, strived, and thrived. Sometimes strong and weak forms live

side by side in a person’s mind, forming doublets like strove and strived or

dove and dived. 27

Doublets usually arise when an irregular form (such as strove) hovers in

a twilight zone in memory and people are not sure whether they have heard

the form or are confusing it with a similar form, like drove. Other doublets

arise for the same reason that you say tomayto and I say tomahto: Britain

and America are divided by a common language. The British prefer dived,

the Americans prefer dove, and people who encounter both dialects, such as

Canadians, are unsure. Often the members of a doublet will diverge in

meaning, grammar, or formality, like twins who strive not to be confused. Shone, for example, is intransitive (without a direct object), as in The stars

shone in the sky, and a touch poetic, whereas shined is an everyday form

that may be used in transitive sentences such as Melvin shined his shoes. (It

would sound silly to say Melvin shone his shoes. ) For many people regular

hanged means “suspended by the neck until dead,” irregular hung merely

“suspended.” Sometimes a muzzy participle will enjoy full vigor as an

adjective, often with its own meaning. For example, smitten is doing fine as

an adjective that means “infatuated,” not literally “walloped” (though the

original metaphor is clear enough, and visible in related metaphors such as

stunning and lovestruck).

Some of the past-tense forms originally in this class became muzzier

and muzzier until they faded out entirely and their verbs became regular. Abode used to be the past tense of abide and today survives only as a noun

meaning “residence.” No speaker of modern standard English uses chide—

chode, glide—glode, gripe—grope, or writhe—wrothe, though some

examples, such as climb—clomb, cling to life in rural areas of Britain and

America. Many of the wayward verbs did not fall into the arms of regularity

but were attracted to other irregular patterns. For example, the short vowel ĭ

is common in participles like driven, risen, and written, and in many weak

verbs, and it inspired bit and hid in the standard dialect of English. In

nonstandard dialects we find dim, writ, strid, smit, div, driv, and the forms

immortalized in the Negro spiritual “Joshua fit the battle of Jericho” and in

the doggerel “Spring has sprung / The grass is ris / I wonder where the

boidies is.” 28

The pairing of ī and ō in rise—rose, drive—drove, and other

descendants of Class I can be seen, with variations, throughout the strong

verbs, where ī-like vowels are frequently replaced by ō-like vowels:

find—found; also bind, grind, wind

freeze—froze; also speak, bespeak, steal, heave, weave

wear—wore; also bear, forbear, swear, forswear, tear

take—took; also mistake, partake, forsake, shake

wake—woke; also awake, break

Forsook and hove are pretty recherché these days, with hove appearing mainly in nautical contexts such as The ship hove to; other uses, such as

Irving hove his lunch, could only be said in jest. Like the other strong

classes, the swear-swore class used to embrace more verbs, but many

defected to the regular side:

But unburied whiten the bones of the crew;

Ah!

would that the widow and orphan but knew

The place where their dirge by deep billows is sighed,

The place where unheeded, unholpen, they died. 29

Some of the old irregular forms survive in rural dialects, such as help-holp,

tell-tole, melt-molt, and swell—swole, and others survive in adjectives in the

standard dialect such as molten and swollen. If you shorten both vowels of the e—o pattern you get:

get—got; also forget, beget, tread

which also beget some muzziness. The participle has got is British, has

gotten American. As with many differences between the dialects, it was the

Mother Country that corrupted the mother tongue; gotten was the form used

in England when the first colonists left in the seventeenth century, and the

Americans preserved it while it vanished in the British Isles. Trod and trodden sound vaguely Winnie-the-Poohish to American ears, because

Americans seldom use the verb to tread: Where the British say tread on,

Americans say step on (notwithstanding one of the slogans of the American

Revolutionary War, “Don’t Tread on Me”).

When tread is used, it is regular:

He treaded water; not He trod water. Begot suffers because of the

familiarity of begat in the King James Bible and the countless satires based

on it. Strangely enough, three common verbs undergo these vowel changes in

reverse:

come—came; also become, overcome (compare wake—woke, take

—took)

fall—fell; also befall (compare get-got)

hold—held; also behold (compare swear—swore)

Came came from a very old irregular whose origins are obscure, but hold

(and maybe fall) really did get reversed. Originally to hold was to held (actually, heal-dan) with past tense hold ( heold).

Similarly, fall used to have the forms feallan—feoll. Some ancient, influential, and confused group of

speakers managed to mix up these verbs with their past-tense forms. This is

not as addled as it may seem; today people occasionally confuse the parts of

a verb when the past tense or participle is more commonly used than the

stem:

Even as environmentalists speak of a seamless web of life, and the

artery advocates speak of a seamless city, the designs on the

drawing board still rent the land from the sea and undermine its

urbanity \[from rend—rent].

30

The videophone is the same size as a regular phone but includes a

3.3 inch color screen with a tiny camera and lens. The...

company

hopes to smitten prospective buyers by renting the phones for less

than $30 a day \[from smite—smote—smitten].31

REEBOK KICKS ITSELF OVER NAME WITH BAD FIT...

For

a company that made its reputation by helping to shod the women’s

aerobics movement, the Incubus name would definitely seem out

\[from shoe—shod; Reebok had named a women’s running shoe

Incubus, not realizing that the word refers to an evil spirit that has

sex with women while they are asleep].

32

Producer Harvey Weinstein hoves his boorish bulk up to the mike

for his moment in the sun for the callow “Shakespeare in Love”—

but is miraculously sent packing by the deus ex machina of the

orchestra \[from heave-hove].

33

Similarly, hoist was originally the past tense and participle of hoise (as in

For ’tis the sport to have the enginer Hoist with his own petar, from

Hamlet), but it has since been reanalyzed as the stem in hoist—hoisted. The following family is a freeze-frame of the process by which neat

classes can get messier over the centuries:

blow—blew; also grow, know, throw, draw, withdraw, fly, slay

What do they have in common?

All end with a vowel, and all begin with a

cluster of consonants except know. In fact even know begins with a

consonant cluster in its spelling, and that tells a story. Spellings usually

reflect old pronunciations, and the k in know was originally spoken aloud;

the word was pronounced k’nawa. So these verbs used to be completely

consistent. Owing to the disappearance of kn and gn at the beginning of

spoken English words, one member no longer fit the membership

requirements and had to be kept in the class by sheer stipulation. In the

history of languages many law-abiding classes become more and more

ragged as general pronunciation shifts mangle their members, until

eventually the criteria become indiscernible to children and the words are

memorized individually. With only one nonconformist member thus far, the blow class has not

yet disappeared, though it has suffered losses. Slay—slew has a biblical feel

and may be on the way out, if we are to judge by recent usages such as Bun

slayed Alexander Hamilton in a duel.34 Crow—crew survives in the bookworm expression The cock crew; even small changes in the expression,

such as The rooster crew, sounds peculiar, and Harvey crew over his victory

is unintelligible. Regional dialects have added or preserved a few more,

such as show—shew, saw—sew, sow—sew, and snow—snew; in 1942 the

Chicago Sun wrote of the weather, It blew and snew and then it thew. These

forms are rarely heard today, however, and the trend is in the opposite

direction: attrition into the regular class. Children make errors such as

blowed and knowed more often than for any other kind of irregular verb. 35

The journalist H. L. Mencken was an assiduous student of the vernacular

speech of the United States and documented many common nonstandard

past-tense forms in his magisterial volumes The American Language. Among them are blowed, knowed, throwed, drawed, and one made famous

by a character in Harriet Beecher Stowe’s Uncle Tom’s Cabin. Theodore

Bernstein, in The Careful Writer, comments on her oft-quoted words:

TOPSY

“In the absence of such reorganization, the city’s court structure as a

whole has just ‘growed,’ like Topsy”; “Like Topsy, that

Government-held surplus of farm commodities just keeps growin’.’”

Once and for all, Topsy’s exact words, punctuated variously in

different editions and in different books of quotations, were: “I

’spect I grow’d.” No “just,” no “jes’,” no “growin’,” no nuffin’.

Anyway, Topsy, Queen of the Clichés, should drop dead. See

Clichés. 36

A few verbs besides came take an ā in the past tense:

eat-ate; also give, forgive, bid, forbid, lie

Bade is a somewhat stilted past tense form of bid in the sense of “ask” or

“command to,” though not in the sense of poker, bridge, or defense

contracts—no one says He hade three clubs. Lie—lay is a trap seemingly

designed to lure speakers into errors and to provide material for the

lamentations of language lovers (including me, in private moments).

A

recent article by Cullen Murphy in the Atlantic Monthly, “The Lay of the

Language,” was devoted to the verb,37 and even the Muppets have been

dragged into the controversy. In 1999 the talking doll Sing and Snore Ernie

had to be reprogrammed after purists objected to his statement, “It feels

good to lay down” 38 (the biggest hooha over a talking doll since Barbie set

back the cause of gender equality by whining, “Math is hard”).

What’s

wrong with lay?

Officially, it belongs to two verbs. One is an intransitive

irregular verb, lie—lay—lain, meaning “recline”:

Stem: Please lie down and tell me about your childhood. Past tense: He lay down on the couch. Participle: He has lain down on the couch. The other is a transitive regular verb, lay—laid—laid, meaning “set down”:

Stem: Lay your cards on the table. Past tense: He laid his cards on the table. Participle: He has laid his cards on the table. Like Ernie, many casual speakers use lay for both—as in I’m going to lay

down—and who can blame them?

As if the sharing of lay in the two

conjugations weren’t confusing enough, the two verbs ought to be one,

according to the grammatical logic of English. Lay means “cause to lie,”

and is one of a handful of verbs meaning “cause to X” that differ by a vowel

from a related verb meaning “to X.” The others are sit—set, rise—raise, fall

—fell (as in to fell a tree), and believe it or not, drink—drench. In most other cases, the verb that means “to X” and the verb that means “cause to

X” sound the same:

The stick leaned against the house. I leaned the stick against the house. The planter stood on the deck. I stood the planter on the deck. The baby sat on the bed. I sat the baby on the bed. The “ungrammatical” intransitive lay follows the pattern of lean, stand, and

sit perfectly. Many purists believe that intransitive lay is a recent corruption,

disseminated by rock lyrics such as Bob Dylan’s Lay Lady Lay and Eric

Clapton’s Lay Down, Sally. But a rule of thumb in language is that any so-

called corruption that occurs frequently enough for the guardians to notice it

will turn out to have been common in the language for a century or more. Intransitive lay was unexceptionable in the seventeenth, eighteenth, and

nineteenth centuries; for example, in 1812 Byron wrote, “There let him lay”

in “Childe Harold’s Pilgrimage.” 39 The historical linguists Thomas Pyles

and John Algeo report:

The brothers H. W. and F. G. Fowler (1931, p. 49) cite with

apparently delighted disapproval “I suspected him of having laid in

wait for the purpose” from the writing of Richard Grant White, the

eminent nineteenth-century American purist—for purists love above

all to catch other purists in some supposed sin against English

grammar.” 40

Another long-term trend reshaping the English language is most apparent in

our final class of irregular verbs, illustrated in a greeting card by Suzy

Becker:

Reprinted with permission of the artist. The ing—ang—ung pattern often is generalized in dialects and in

affectations of dialects, as in the jocular Who would have thunk?

In 1998

the Texan columnist Molly Ivins entitled a book You’ve Got to Dance with

Them What Brung You, allegedly a backwoods aphorism though more likely

an urbanite’s attempt at hick-chic. The baseball pitcher and sportscaster

Dizzy Dean was said to have narrated a play as follows:

The pitcher wound up and flang the ball at the batter. The batter

swang and missed. The pitcher flang the ball again and this time the

batter connected. He hit a high fly right to the center fielder. The

center fielder was all set to catch the ball, but at the last minute his

eyes were blound by the sun and he dropped it!

41

Dave Barry, defending himself against enraged Neil Diamond fans after

making a joke at the singer’s expense in a prior column, describes the

results of a reader survey:

Unfortunately, a lot of survey voters are not so crazy about Neil’s

work, especially the part of “Play Me” where he sings, “...

song she

sang to me, song she brang to me...” Of course I think those lyrics

are brilliant; however, they brang out a lot of hostility in the readers. The ing—ang—ung pattern came down to us from another class of strong

verbs in Old English, Class III, which included singan—sang—sungen. Many modern verbs follow it to varying degrees:

ring—rang—rung; also sing, spring, drink, shrink, sink, stink,

swim, begin

cling—clung; also fling, sling, sting, string, swing, wring, slink,

stick, dig, spin, win

run—ran—run

hang—hung

strike—struck

sneak—snuck

sit—sat, spit—spat

Most of the ing—ang—ung verbs end in -ing or -ink. Two of the others deserve comment. Begin has the distinction of being the only common irregular verb that

is neither monosyllabic nor built around a monosyllabic root. (The

common, Anglo-Saxon words we use every day tend to be monosyllables,

and the irregular verbs are no exception.) Begin is formed with the prefix

be; as in the similar irregulars become, befall, beget, behold, beset, and

bespeak. In begin s case, however, the residue, -gin, is not an English word;

it came from a now-defunct Proto-Germanic verb meaning “open.” (There

are two other irregular past-tense forms, both somewhat unusual, whose

stems cannot stand alone as verbs: forsake—forsook and beseech-besought.)

Snuck has the distinction of being the most recent irregular to enter the

standard language, with a first citation in the Oxford English Dictionary

from 1887.

According to a recent survey, most younger Americans have no

problem with snuck, though most older Americans frown on it.42 William

Safire quotes a letter from Doris Asmundsson, a professor emerita of

English: “Words like creak, critique, eke, freak, leak, and tweak do not, in

the past tense, become cruck, crituck, uck, fruck, luck, and twuck. Why then

snuck?

Eventually a sneaker might turn into a snucker.” 43 According to one theory, snuck sneaked into English via sound symbolism. Its

connotation of quickness, furtiveness, and mild disreputability brought to

mind the sound pattern of slunk and suck, especially since all three end in a

suitably crisp k. 44 A less far-fetched explanation is that sneak is close in pronunciation to sting, strike, dig, and especially stick—an ĭ is just a lax, short ē, and n is basically t or d pronounced through the nose, as any cold-sufferer can tell you. The failure to rhyme with creak and tweak was no

impediment, because similarity in the gestures of articulation matter more

than similarity in sound, and that makes it tempting to analogize stick—

stuck to sneak—snuck. Many dialectal past-tense forms that don’t rhyme exactly with cling or

slink still take the ŭ vowel in the past tense. Mencken and others report climb—clumb, shake—shuck, take—tuck, dive—duv, and drive—druv, also

heard in the English proverb “Sussex won’t be druv.” One speaker

described what they used to do to endangered species in the olden days as

follows: They killed ’em and skun ’em out. Dizzy Dean was famous for

saying He slud into second, and some baseball fans say, “If Dykstra hadn’t

dropped the ball, the runner wouldn’t have tug (tagged).

45 On the following

page is another common example. The ing-ang-ung verbs are a bellwether of a millennium-old and still

ongoing trend in the English language. In the fourteenth century the

egalitarian preacher John Ball roused the rabble with the slogan “When

Adam delved and Eve span / Who was then a gentleman?” Span was the

past tense of spin, following the i—a—u pattern of verbs such as sing, swim,

and begin. But eventually the participle spun usurped the past-tense slot,

relegating span to the dustbin of history. These takeovers are still going on. When the Walt Disney corporation released a film called Honey, I Shrunk

the Kids, English teachers were up in arms: It should be Honey, I Shrank the

Kids, they said. Nonetheless, most people say shrunk, sprung, sunk, and

stunk, not shrank, sprang, sank, and stank. (Some go the other way: At an

infamous moment in the O. J. Simpson murder trial, the prosecutor

Christopher Darden said hopefully, “The gloves appear to have shrank

somewhat.”)

ARLO \& JANIS® by Jimmy Johnson

ARLO \& JANIS reprinted by permission of Newspaper Enterprise Association, Inc. The teachers are fighting a losing battle because even the language

mavens are losing their grip on the distinction. William Safire got an earful

from the Gotcha!

Gang and the Uofallpeople Club when he wrote,

“Trivialize had its moment in the vogue-verb sun, until the usage of this

older verb shrunk to the very occasional.” 46 The Boston Globe’s language

maven, Jan Freeman, wrote that she once did a double-take upon hearing

They sort of sprang it on me, momentarily thinking it was incorrect.47

Shrank, together with the other ank and ang words, is under assault from two directions—from its own past participle shrunk, and from the

many ing verbs that have already lost their angs and really do take the ung form in the past tense as well as in the participle: He slung (not slang) the

hash; They strung (not strang) him up with a rope, He flung (not flang) the ball at the batter. Surely and steadily, ing-ung-ung is displacing ing-ang-ung, part of a larger erosion of the distinction between participles and past

tense forms throughout the verb system. Regular verbs fail to distinguish pasts and participles at all —I walk, I

walked, I have walked—and fewer than half of our irregular verbs continue

to distinguish them; most are like mean—meant—has meant or find—found

—has found. In nonstandard dialects the distinction is even feebler. I seen it

and A man come into the bar are absolutely standard outside the upper and

middle classes, even in urban areas, and He begun to cry, She done it, and

They gone home are also common. In the early decades of the twentieth

century, Mencken reported the past-tense forms div, driv, riz, swole, taken,

thrown, and writ, and the participle forms ( has) ate, blew, broke, did, drank,

drive, froze, gave, rode, rose, ran, stole, swam, took, tore, woke, wore, and

wrote. Nonetheless, many people assume that the erosion is a recent

development:

Dear Ann Landers:

Have Americans forgotten there is such a thing as verb tense?

I am

shocked when I hear people say “woulda came,” “coulda went,”

“shoulda did,” “woulda took,” “had went,” “hadn’t came,” and so

on. Don’t they realize “woulda” and “coulda” are slang versions of

“would’ve” and “could’ve”—which are contractions for “would

have” and “could have”?

I heard a narrator say, “I seen” in a political commercial, and a

TV reporter say, “We haven’t spoke.”...

A TV anchorwoman said,

“had threw it” and “between you and I.”

I am a secretary for almost 50 years and am thankful that, with

only a high-school education, my English is impeccable. You will

do a lot of folks a big favor if you print this letter and bring it to

their attention. E. E. Wood Ridge, N. J. Dear E. E.:

Thanks for taking the time and trouble to write. I shoulda thunk to

tell them off myself. 48

Confusions of past and participle forms are easy to explain. Some may

originate in mishearings. As E. E. pointed out, the auxiliaries has and have

that signal the perfect construction are often contracted to he’s, we’ve,

could’ve, should’ve, and would’ve, or even coulda, shoulda, and woulda. (Anyone who has graded student papers or dipped into internet discussion

groups is also familiar with could of, should of, and would of. ) That makes

the haves easy to miss in rapid speech; He’s seen it, in particular, is easily

reanalyzed as He seen it. Yet the main reason for the decline of the ang—ung distinction is that

all distinctions in English inflection have been declining for the past

thousand years; syntax has been shouldering the load formerly borne by

morphology. Old English and Middle English had separate verb forms not

only for present, past, and participle, but also for different persons ( I, you,

he/she) and numbers (singular and plural) within the past tense. The past

forms for sing, for example, would have been:

I sang

We sungon

Thou sunge

You sungon

He/she sang

They sungon

When the person and number distinctions collapsed, every verb had to end

up with a single past-tense form, and a game of musical chairs broke out,

with the different stems competing for the remaining chair. With some

verbs the singular won, such as sing—sang—sung; with others the plural

won, such as sling—slung—slung (the past plural was usually similar to the

participle—the phenomenon of syncretism we met in chapter 2).

Another

free-for-all took place among the participles of the verbs that kept their -en

and had to grab a stem from the collapsing conjugation. Some took the stem

of the base form, such as take—took—taken, others took the stem of the past

form, such as break—broke—broken, and still others kept their own stem,

such as swell—swelled—swollen. Some participles can jump ship to another

pattern: If you apply the break—broke—broken pattern to shake and take,

you get the somewhat cutesy shooken and tooken. I have been advised that

tooken has become standard in Generation X circles, but if it is, do not

blame it on their ethos of ironic detachment; it was used as early as 1946 in

“Put That Kiss Back Where You Found It,” a song recorded by Benny

Goodman: “Took it when I wasn’t lookin’/ And my heart you’ve also

tooken.”49 The steady erosion of distinctions in English inflection helps us

understand why we continue to be confused by verbs such as shrink and

spring in the second millennium after the end of Old English. For the sake of completeness, here are the remaining irregular forms. Shorn and swollen belong to a small group of verbs that are regular except

for their participles:

swell—swelled—swollen, shear—sheared—shorn

show—showed—shown; also sow, sew, prove, strew

(A few other irregular participles were orphaned from their verbs and

survive only as adjectives, most of them somewhat unusual: bereft,

unbidden, clad, cleft, cloven, drunken, forlorn, girt, gilt, misbegotten, hewn,

beholden, laden, molten, mown, pent, misshapen, clean-shaven, shod,

sodden. ) I couldn’t figure out where to put these:

beat—beat—beaten

choose—chose—chosen

see—saw—seen

stand—stood; also understand, withstand

The stand—stood pattern is heard in the oft-cited plaint of the fight

promoter Joe Jacobs, “I should have stood in bed,” and in the modal

auxiliary verb

can—could

which retains a present-past contrast in usages like I can’t polka now, but I

could before I broke my leg. Other pairs of modal auxiliaries— may—might,

will—would, and shall—should—began life as different tenses of the same

verb, but the couples divorced long ago and might, would, and should are

no longer past-tense forms. Exactly how many irregular verbs are there in the English language

today?

If we don’t double-count prefixed families such as get and forget,

don’t count dialectal form such as drug and brung, do count verbs that are

irregular either in standard American or standard British English, and do

count the muzzy but widely recognizable forms, we end up with 164

modern irregular verbs: 81 weak (ending in t or d), 83 strong. Compare this

to Old English, with 325 strong verbs alone, and it is clear that English is

indeed becoming “weaker.” In later chapters we will see whether the

surviving but endangered irregular verbs are sustainable. We have seen how the weak past-tense forms can be traced to Proto-

Germanic about 2000 years ago, and the strong forms can be traced back to

Proto-Indo-European at least 5500 years ago. But where did they come

from?

They certainly were not designed by a committee, and presumably

did not arise from divine revelation. No one knows the answer, but a few

brave linguists have speculated. The dental suffix in Proto-Germanic, the ancestor of our -ed, may have

come from a reduced form of the verb to do. 50 Many languages use an empty verb like do as an auxiliary verb that carries information about the

statement as a whole, such as tense, degree of completion, and negation. Indeed, Modern English uses do for that purpose in yes-no questions ( Do you want to dance?

) and in negations ( Alice doesn’t live here anymore).

In

the history of a language, prefixes and suffixes often arise from the erosion

of verbs such as do, take, be, and have, a process called

grammaticalization. 51

If the dental suffix came from do, it would explain why it has the sound

d or t. In Proto-Germanic, do could come after a noun or another verb, very

roughly like He hammer-did or She walk-did. The do could have eroded to

the stub d and attached itself to the verb, giving us the ancestor of -ed. The theory also explains why -ed has become the regular suffix,

applying freely to any new or strange verb. The phrase containing do and a

verb would have been created by the rules of syntax, the combinatorial

system par excellence, which allows almost anything to combine with

anything else. A promiscuous auxiliary verb would have been a natural

ancestor of a promiscuous suffix: Just as a verb like do can combine with

any verb at all ( He did abandon, He did abate, He did abbreviate, and so

on), so its descendant -ed might have retained this habit, allowing it to

combine with any verb at all ( abandoned, abated, abbreviated, and so on).

The Indo-European ablaut or vowel-change patterns, the ancestors of

our strong verb forms, change an e (a sound between Ed and aid) or a neutral vowel to a (as in father), or to ō (as in hoe or horse).

The e is pronounced with the tongue hump toward the front of the mouth, the a and

ō with the tongue low and toward the back. This contrast, between a higher

front vowel and a lower back vowel, survives in the majority of modern

English irregular verbs. The base forms have sounds like ĕ and ē and ĭ and ī

and ā, and the past-tense forms have sounds like ă and ŏ and ō and ŭ and. That may not be a coincidence. Three of the great linguists of the

middle decades of the twentieth century, Roman Jakobson, Jerzy

Kuryłowicz, and Morris Swadesh, noticed that in many languages the

vowels pronounced with the tongue high and at the front of the mouth tend

to be used for the basic forms of nouns and verbs (such as the singular form

of a noun and the infinitive of a verb), whereas the vowels pronounced with

the tongue lower and farther back tend to be used for the specially marked

forms (such as plural nouns and tensed verbs).

52 Moreover, the higher and

farther front vowels have different connotations from the lower and farther

back vowels in pairs of contrasting words. The high front vowels come first

in expressions such as pitter-patter and dribs and drabs; we don’t say

patter-pitter or drabs and dribs. And in pairs such as this and that, here and there, and me and you, the higher and farther-to-the-front vowels are found

in the word that means “self” or “near the self,” the lower and farther-to-

the-back word means “other” or “far from the self.” That is true not only in

English but in many families of languages.53

Perhaps this ubiquitous vowel contrast is a case of sound symbolism. The linguist Roger Wescott has pointed out that high front vowels are

pronounced with a constricted mouth cavity and the tongue close to the

visible part of the vocal tract, whereas low back and central vowels are

pronounced with a large mouth cavity and the tongue buried from view. That may call to mind the conceptual distinction between presentness and

pastness. Pastness may remind people of a cavity or space, because a past

event is separated by an interval of time from the present moment, and

metaphorically speaking time equals space. It may also remind people of

remoteness or distance, because metaphorically speaking long ago equals

far away. Perhaps as Indo-European was developing, speakers vaguely felt

that lower and farther back vowels fit better with the concept of an event

separated in time from the present, and that higher and farther front vowels

fit better with an event in the here and now.54 Of course, the Indo-

Europeans had to pick some vowel contrast if they were to mark tense with

a vowel, and for all we know they could just have easily gone the other

way. But the fact that the vowel contrast appears in many unrelated

languages with similar roles, and was preserved and embellished in our own

5500-year game of Broken Telephone, hints that it might have some

semantic resonance for human minds. * For a chart that summarizes the history, dates, and family affinities of the English language, see

page 212.

4

IN SINGLE COMBAT

Now that we know all about regular and irregular verbs, how well does

the words-and-rules theory hold up?

In some ways quite well; in

other ways not so well. According to the theory, the ingredients of language

are a list of memorized words, each an arbitrary pairing between a sound

and a meaning, and a set of productive rules that assemble words into

combinations. Regular and irregular forms exemplify the two ingredients:

Regular forms are generated by rule, irregular forms are memorized by rote. The dissection of language in chapter 2 showed that the organs that

secrete regular forms are elegant combinatorial systems, just as the theory

had led us to expect. But the excavation in chapter 3 showed that the

depository for irregular verbs is not disorganized and inert, which is what

the theory had led us to expect. The irregular verbs are shot through with

patterns:

blow—blew, grow—grew, know—knew, throw—threw

bind—bound, find—found, grind—ground, wind—wound

drink—drank, shrink—shrank, sink—sank, stink—stank

bear—bore, swear—swore, tear—tore, wear—wore

This is not what we would expect if they were a laundry list of arbitrary items. Suppletive pairs such as go—went and be—was are the exception,

whereas if the irregulars were truly acquired one by one they could just as

easily have been the rule. We have seen that many irregular patterns are fossils of extinct rules

that lived in the heads of speakers long ago, but history can explain only

part of the patterning. If the irregulars got all of their patterning from old

rules and have degenerated steadily since the rules became extinct, today’s

patterns should be tattered versions of the ancient Indo-European strong

classes, with a core of surviving verbs that fit the old patterns and a halo of

distorted verbs that have drifted off in various directions. But many verbs

have joined the strong classes, or have jumped from class to class, and

today’s irregular families cross-classify the original ones. Ring—rang

originally was weak (with a past tense like ringed) and was attracted to the

ing—ang—ung class by analogy to verbs like sing. The same happened to

dig—dug, stick—stuck, wear—wore, show—shown, and many others. Some

verbs that entered the language after the Old English period also were

seduced into strong classes, such as fling—flung and sling—slung. Others

switched from one strong class to another, such as slay—slew and draw—

drew (originally drough).

Still others found the weak irregular patterns appealing, such as light-lit and creep-crept.1

The words-and-rules theory would be off the hook if these attractions

and conversions were a thing of the distant past. Perhaps the vowel-change

rules did not die out completely with the Indo-European or Germanic tribes,

but lingered in weakened form in parts of England for a few centuries

before giving up the ghost. Unfortunately, some of the conversions are

fairly recent. Kneel—knelt, dive—dove, catch—caught, and quit—quit

became popular only in the nineteenth century; George Washington, for

example, used catched, and Jane Austen used quitted. And as we saw, snuck

came into English a century ago and is only now becoming standard. The funny irregulars in the nonstandard dialects of English add to the

worries. While some rural irregulars are quaint holdovers of old standard

forms, such as help—holp, climb—clim, and creep—crope, many, if not

most, are homegrown products of the creativity of local speakers: bring—

brang—brung, dive—div, chide—chode, snow—snew, climb—clomb, drag—

drug, slide—slud, fling—flang, and literally hundreds of others.2

As an experimental psychologist I have been trained not to believe

anything unless it can be demonstrated in the laboratory on rats or

sophomores. To my knowledge no one has yet studied irregular verbs in

rats, but the linguists Joan Bybee and Carol Moder have studied them in

sophomores and have shown that they are all too happy to generalize

irregular patterns to novel verbs. They asked the students in a university

linguistics course (not only sophomores, of course) to write down the past-

tense forms of existing and made-up verbs, completing sentences such as

Sam likes to spling. Yesterday he \_\_\_\_\_\_\_\_.

Almost 80 percent of the

subjects offered splang or splung. Even when given real words, some of the

subjects were tempted by irregular patterns and came up with creative

forms such as dig—dag, sting—stang, slink—slank, streak—strack, skid—

skud, and clip—clap. Bybee and Moder may have duplicated in the

laboratory the process that gave us forms lik e fling—flung and drag—drug

in the history of English. 3

But maybe not. University students treat anything that looks like a test

as a test, and perhaps they saw the task as a challenge to their ingenuity in

coming up with creative forms. We know that outside the lab, people

ruminate over irregular patterns and sometimes deduce ways in which they

ought to be generalized. Here is Leo Rosten and Leonard Ross’s character

Hyman Kaplan, a Jewish immigrant learning English at night school:

“I tought de pest time ‘bite’ should be—‘bote.’”

Miss Mitnick gave a little gasp. “‘Bote’?” Mr. Parkhill asked in amazement. “‘Bote’?”

“‘Bote’!” said Mr. Kaplan. Mr. Parkhill shook his head. “I don’t see your point.”

“Veil,” sighed Mr. Kaplan, with a modest shrug, “if is ‘write,

wrote, written’ so vy isn’t ’bite, bote, bitten’?”

Psychic cymbals crashed in Mr. Parkhill’s ears. “There is not such a word ‘bote,’” protested Miss Mitnick, who

took this all as a personal affront. Her voice was small, but

desperate. “Not-soch-a-void!” Mr. Kaplan repeated ironically. “Mine dear

Mitnick, don’ I know is not soch a void?

Did I say is soch a void?

All I’m eskink is, isn’t logical should be soch a void?”4

The students in the spling experiment also may have thought to themselves,

“Isn’t logical should be soch a void?,” and may have concocted forms they

never would have used when speaking naturally. A related worry is that people make note of irregular forms, especially

in other people’s dialects, and consciously use them in wit and wordplay, as

we discovered in the previous chapter and can see in these old jokes:

A woman gets into a taxi in Boston’s Logan airport and asks the

driver, “Can you take me someplace where I can get scrod?” He

says, “Gee, that’s the first time I’ve heard it in the pluperfect

subjunctive.”

A friend of mine came across some cut flowers that were so

spectacularly red she thought they must be fake. “These are

amazing,” she said. “Are they dyed?” The florist shook his head. “No, no, not at all,” he said. “Just put ’em in water and they’ll be

fine.” 5

A man was on trial for pulling a woman down the street by her hair. The judge asked the arresting officer, “Was she drugged?” The

policeman replied, “Yes, sir, a whole block.” 6

Verbal humor depends on the audience’s recognizing that an odd word is

odd; if the words were unexceptional, there would be no joke. So just

because some wiseguy finds a novel past-tense form logical or amusing, it

does not mean that the form is a natural product of the system he uses in

everyday speech. Instead the form may be a product of one’s intellectual

faculties reflecting back onto one’s language, an ability called

metalinguistic awareness. But one kind of generalization cannot be a product of conscious

cogitation, namely, the errors preschool children make in their spontaneous

speech:

It was neat—you should have sawn it!

Doggie bat me \[bit].

The cheerios got aten by the Marky. I know how to do that. I truck myself \[tricked].

He could have brang his socks and shoes down quick. And they swang into a roller coaster and we went with their cars

and they were sliding and they did a leap. Elsa could have been shotten by the hunter, right?

So I took his coat and I shuck it \[shook].

This is the best place I ever sot \[sat].

I bate Paul up \[beat].

You mean just a little bitty bit is dranken?7

The psychologist Fei Xu and I combed through transcripts of the speech of

nine children in an electronic archive and pulled out all the past tense and

participle forms, 20,000 in all.8 We found numerous irregular forms that are

not standard English:

Children don’t make these errors very often—only in two tenths of 1

percent of the opportunities—but eight of the nine children made at least

one while the tape was running, and we know that children continue to

make them well into the school-age years. I have a drawing from a seven-

year-old girl with the caption Win I Wit to Git a crismis chre and it wus

Sowing so i braing my umbrella 9 and I remember that when I was twelve I

had a persistent urge to use raught as a past tense of reach, on the analogy

of teach—taught. The irregular patterns refuse to die. Irregular verbs are supposed to be a list

of arbitrary words memorized by rote, just like duck and walk, with only a

trace of patterning left behind by long-defunct rules. Instead, people extract

the patterns and extend them to new words, just as they do with the regular

pattern in errors like breaked, in neologisms like moshed, and in the wug-

test. The distinction between regular and irregular inflection, and therefore

between words and rules, is not so clear anymore. Either the irregular

patterns are generated by rules, just like the regular pattern, or linguistic

productivity does not depend on rules in the first place but can arise from

words via some ability to associate the patterns in known words with the

patterns in new ones. Both alternatives have been developed into famous, full-blown theories

of the English past-tense system. Their clash is one of the most vigorous

controversies in the modern study of the mind, echoing through psychology,

linguistics, philosophy, computer science, and neuroscience. According to

the theory of generative phonology developed by Noam Chomsky and

Morris Halle, rules rule.10 Every drop of patterning in past-tense forms, regular or irregular, is squeezed out into rules, and only the compressed,

desiccated residue is stored in the mental lexicon. According to the theory

of parallel distributed processing or connectionism developed by David

Rumelhart and James McClelland, there are no rules: People store

associations between the sounds of stems and the sounds of past-tense

forms, and generalize the associations to new words if they are similar to

old words. 11 Both theories invoke a single kind of mental computation to

explain how people generate regular and irregular forms, but for generative

phonology it’s rules all the way down, whereas for connectionism it’s

memory all the way up. A clash over irregular verbs may call to mind the remark that academic

debates are heated because so little is at stake. But in this case something is

at stake. The past-tense debate is the latest battle in a centuries-old

disagreement over two very different ways of understanding the mind:

When a man reasons, he does nothing else but conceive a sum total

from addition of parcels, or conceive a remainder from subtraction

of one sum from another; which, if it be done by words, is

conceiving of the consequence of the names of all the parts to the

name of the whole, or from the names of the whole and one part to

the name of the other part....

For REASON is nothing but

reckoning. 12

In this passage from Leviathan, written in 1651, Hobbes uses reckoning

in the original sense of counting, calculating, or computing. For example,

suppose the definition of “man” is “rational animal.” Then if we are told

that something is “rational” and an “animal” (names of parts) we can

deduce it is a “man” (name of whole), and if we are told that something is a

“man” (name of whole) and that it is “rational” (name of one part) we can

deduce that it is a rational “animal” (name of the other part).

These steps

could be laid out as mechanical instructions to recognize and copy words, a

kind of symbol, and therefore could be “reckoned” or computed by

someone who has no idea what the concepts “rational” and “animal” even

mean. If the symbols are patterns in the brain rather than words on a page,

and the patterns trigger other patterns because of the way the brain is wired,

then we have a theory of thinking. Among the people influenced by Hobbes was Leibniz, who was inspired

as well by John Wilkins and other designers of artificial languages

discussed in chapter 1.

Leibniz took Hobbes literally when he said that

reason is nothing but reckoning. He devoted much of his life to inventing a

scheme that would perfect the computations underlying thought, turning

arguments into calculations and making fallacies as obvious as errors in

arithmetic. “Once this has been done,” he wrote, “if ever further

controversies should arise, there should be no more reason for disputes

between two philosophers than between two calculators. All that will be

necessary is that, pen in hand, they sit down together at a table and say to

each other...

‘let us calculate.’”13 In one version of Leibniz s scheme,

“man” is assigned the number 6, “animal” is assigned 2, and “rational” is

assigned 3.

Since 2 x 3 = 6, a rational animal must be a man; since 6 -s- 3 =

2, a man is not just any old rational being but specifically a rational animal. If the number for “monkey” is 10, one may calculate that monkeys are not

men or vice versa, and that monkeys, while animals, are not rational. 14

The idea that intelligence arises from the manipulation of symbols by

rules is a major doctrine of the school of thought called rationalism,

generally associated with Leibniz and Descartes. When the symbols stand

for words and the rules arrange them into phrases and sentences, we have

grammar, the subject of Cartesian linguistics, which later inspired

Humboldt and then Chomsky. When the symbols stand for concepts and the

rules string them into chains of inference, we have logic, which became the

basis for digital computers, the artificial intelligence systems that run on

them, and many models of human cognition. 15

But symbol manipulation is not the only way the mind might work:

There appear to be only three principles of connection among ideas,

namely, resemblance, contiguity in time or place, and cause or

effect. Experience shows us a number of uniform effects, resulting

from certain objects. When a new object, endowed with similar

sensible qualities, is produced, we expect similar powers and forces,

and look for a like effect. From a body of like color and consistence

with bread we expect like nourishment and support.16

In this passage from his 1748 Enquiry Concerning Human

Understanding, David Hume summarizes the theory of associationism, a

major tenet of the school of thought called empiricism. The mind connects

things that are experienced together or that look alike (Hume later

eliminated cause and effect as a separate principle) and generalizes to new

objects according to their resemblance to known ones. Just as the

rationalists were obsessed by combinatorial grammar, the associationists

were obsessed by memorized words. In his 1689 Essay Concerning Human

Understanding, John Locke pointed to the arbitrary connection between

words and things as the quintessential example of how the mind forms

associations by contiguity in time: We learn dog when Mother says “dog” in

the presence of a dog. Replace Locke and Hume’s “ideas” or “sensible

qualities” with “stimuli” and “responses,” and you get the behaviorism of

Ivan Pavlov, John B. Watson, and B. F. Skinner. Replace the ideas with

“neurons” and the associations with “connections,” and you get the

connectionism of David Rumelhart and James McClelland. The great debate between rationalism and empiricism is familiar to

everyone who has taken a course in philosophy, psychology, or the history

of ideas. It embraces such issues as whether the mind is packed with innate

structure or is a blank slate on which the environment writes, and whether

knowledge comes from making deductions using theories or gathering data

from observation. But the issue that concerns us here is the nature of our

mental machinery, in particular, whether intelligence arises from the

manipulation of symbols or from associations between sensory qualities. And the competing theories of the English past tense provide us with an

unusual opportunity. It’s been three hundred and fifty years since Leviathan, and scholars are

still debating rationalism and empiricism. Both sides appeal to theoretical

coherence, intuitive plausibility, political ramifications, and harmony with

the morals of modern science, but these are pretty squishy criteria, and the

debate rages on. We need a concrete, richly studied instance of human

psychology in which the two grand theories can go head-to-head in

explaining the same facts. In ancient warfare, an army sometimes would send its mightiest warrior

to face his counterpart from the opposing army in single combat. The

outcome would embolden one side in the actual battle that followed, or

might pre-empt it altogether, sparing unnecessary bloodshed. The most

familiar example is the biblical story of David and Goliath, and in The

Right Stuff Tom Wolfe suggested that a contemporary example may be

found in the space race of the 1960s, in which the Mercury astronauts were

treated like single combat warriors against the cosmonauts of the Soviet

Union. Scientific debates sometimes work like single combat, not because

the combat metaphor is particularly apt or appealing, but because it is easier

to compare two great big theories when each is vested in a highly specific

hypothesis and the hypotheses compete on the same ground. The English past tense is the perfect site. The phenomenon is

circumscribed and therefore tractable to study. Its history, acquisition, and

patterns of use have been abundantly documented. It has obvious rulelike

and memorylike features that serve as hurdles that any theory must clear. And it has brought out the best in the ingenuity of contemporary

theoreticians, with each side devising a clever, elegant, detailed, and

surprising model. The past tense is the only case I know in which two great

systems of Western thought may be tested and compared on a single rich set

of data, just like ordinary scientific hypotheses. What are the facts to be explained?

Irregular verbs defy the suggestion that

they are memorized by rote because they show three kinds of patterning. First, irregular past-tense forms are similar in sound to their base forms. For example, drink and drank share d, r, a vowel, n, and k; the only difference is that drank has the vowel a where drink has the vowel i. Similarly similar are swear and swore, sleep and slept, freeze and froze. In

fact, all the irregulars except go—went and be—was share material with

their stems. It didn’t have to be that way. One can imagine a hypothetical

language in which most of the verbs are like go—went, with nothing in

common between stem and past, each stem and each past stuffed into its

own memory slot. We need an explanation of why English does not look

like that. Let’s call this pattern stem-past similarity. Second, a few kinds of change from a stem to its past are seen over and

over among the 164 irregular verbs. The ĭ-ă-ŭ pattern in drink—drank—

drunk, for example, is found, with variations, in sing-sang—sung, sit—sat—

sat, begin—began—begun, shrink—shrank—shrunk, and twenty other

verbs. Similarly, we have freeze—froze, speak—spoke, and steal—stole;

bleed—bled, breed—bred, and feed—fed; teach—taught, fight—fought, and

bring—brought. One can imagine a language in which every verb picked its

own substitution of vowels and consonants from among the thousands that

are logically possible. But generations of learners have passed down an

English language that is very different from that possibility. Let’s call this

pattern, in which the change from stem to past in one verb is similar to the

change from stem to past in another verb, change-change similarity. Third, the verbs undergoing a given irregular change are far more

similar than they have to be. If you are a verb and want to undergo the ĭ-ă-ŭ

pattern, all you really need is an ĭ.

But the verbs that do follow the pattern

( drink, spring, shrink, and so on) have much more in common; most begin

with a consonant cluster like st-, str-, dr-, sl-, or cl-, and most end in -ng or -

nk. Similarly, the verbs whose pasts end in -ew (blow, grow, throw, slay,

draw, and fly) tend to begin with a consonant cluster and end with a vowel. Verbs of a feather change together, and not just in sightings by word

watchers. People extend old patterns to new verbs, as in bring—brang, fight

—fit, and spling—splung, only when the new verb is highly similar to old

ones in memory. We need an explanation of why the human mind is so

impressed by similarity in sound; let’s call this pattern stem-stem similarity. A theory with rules for irregular verbs, as well as regular ones, could

explain all three kinds of patterning. Imagine a rule that said, “If a verb has

the sound consonant-consonant-ĭ-ng, change ĭ to ŭ.

” Notice that the rule doesn’t spell out the past-tense form letter by letter; it just says, “Change

the vowel.” The rest of the input—the consonants before and after the

vowel—come through in the output untouched. We have an explanation for

stem-past similarity. Now suppose that the mind prefers simple grammars, with a few rules,

to complex grammars, with many rules. If there are fewer rules than verbs,

many verbs will have to share a rule, such as “Change ĭ to it. ” We have an

explanation for change-change similarity. Finally, notice that the rule has a condition on it: apply only to verbs

that have two consonants before the vowel and an ng sound after it. The

condition is a gatekeeper that allows in verbs that are similar to the cling-

clung family and filters out those that merely contain ĭ.

This could explain

stem-stem similarity. Now, any drudge can go over the list of irregular verbs in the preceding

chapter and write down a set of tedious rules for them: “If a verb begins in s

and ends in ee, change the ee to aw, ” and so on. But that would be no improvement over the original list of verbs. A theory invoking rules must

be more than a summary of the patterns among the verbs. It should be a

psychological theory: a hypothesis of the format in which children acquire

words, and an explanation of why verbs have the kinds of patterns they do. The trick is to find a compact set of rules that captures the generalizations

the mind likes to make. By far the most ambitious theory of this kind comes from Noam

Chomsky and Morris Halle’s 1968 magnum opus The Sound Pattern of

English, later refined by Halle and the linguist K. P. Mohanan.17 Their rules for irregulars are part of a larger set of rules that capture the sound

pattern of a language (its accent).

Clearly the speakers of a language know

more than the list of words that happen to be in the language at a given moment. For example, English speakers intuit that blicket, dax, and fep are

not English words but could be, whereas ftip, rtut, and nganga are not

English words and could not be (though speakers of other languages might

recognize them as possible words in their languages).

English speakers also

know that when divine is joined to -ity to create divinity, the i vowel in -in-changes from ī to ĭ, and that when Canada is joined to -ian to become Canadian, the final -a vanishes, the stress shifts from the first syllable Ca to the second syllable na, and the vowel in that syllable changes to ā.

Chomsky, Halle, and Mohanan accounted for the patterns of thousands of

English words with just a few dozen phonological rules, each assigned to

one or more of the boxes in the diagram on page 23: lexicon, morphology,

syntax. Their theory comes from a field called generative phonology, a

division of generative linguistics, the approach to language founded by

Chomsky. By positing rules that replace consonants and vowels (phonemes) in the

irregular verbs, Chomsky and Halle enjoy the advantage of rules in general:

accounting for patterns among verbs and changes, and for speakers’ ability

to generalize them. Amazingly, Chomsky, Halle, and Mohanan handled

most of the dizzying patterns among the 165-odd irregular verbs with only

three rules. Virtually all their other rules are needed to explain the sound

pattern of English in general. Chomsky, Halle, and Mohanan roundly reject the words-and-rules

dichotomy. Verbs sit on “a continuum of productivity and generality that

extends from affixation of the -ed suffix in decide—decided to total

suppletion in go—went, ” with families like sing—sang, ring—rang, and

bind—bound, wind—wound in between.18 At one end of the continuum are

the regular verbs, which are handled by a general rule that says nothing

about the words it can apply to. At the other end of the continuum are

suppletive verbs such as go and went, which are simply listed as pairs. In

between are the other irregulars, which are handled by a smaller set of rules,

each tagged to apply to certain verbs. By stipulating which verb may be

touched by which rule, the theorists circumvented the problem of crafting

the rules to single out verbs by their sounds—no small matter, given that

shrink has the past tense shrank, sling has the past tense slung, bring has the past tense brought, and blink has the regular past tense blinked. Another rein on rules keeps each one in a stratum (a component or

subcomponent) so that it does not run wild and apply where it shouldn’t. To

generate keep—kept, for example, Chomsky, Halle, and Mohanan invoked a

rule that shortens a long vowel (changing ē to ĕ) when it occurs before a

consonant cluster, such as pt. But that rule cannot be allowed to apply

across the board or it would turn seeped into sept, wiped into wipped, and

so on. So Halle and Mohanan proposed that the -t and -d found in weak

irregular verbs like kept is not the same as the -ed found in regular verbs,

despite their similar pronunciations. Whereas the regular -ed is attached in

the morphology box on page 23, the -t or -d in the irregular verbs is attached in the lexicon box, which is also where the shortening rule is

confined. This may seem like cheating, but there are independent grounds

for it. Other semiregular sort-of-rules that have nothing to do with the past

tense, like the ones creating serene—serenity and volcano-volcanic, need

the shortening rule too. That supports the idea that several rules are

sequestered together in their own little community. One move that allowed Chomsky, Halle, and Mohanan to get away with

so few rules was factoring apart each complex change into several simple

ones and allowing the simple ones to be assigned in different combinations

to different verbs. For example, tell—told is produced by at least two rules,

one that changes the vowel, the other that adds the -d. The rule that changes

the vowel is also put to work in swear-swore, and the rule that adds -d is

also put to work in flee—fled. Looking at the crisscrossing patterns in tell—

told, swear—swore, flee—fled, bend—bent, burn—burnt, deal—dealt, breed

—bred, and hit—hit, one can readily see that sharing minirules that add t s

and d s, that delete extra ones, and that fiddle with vowels is far more

economical than building a special rule for each family of rhyming verbs. A monumental contribution of generative phonology was to slice rules

even more finely so that they apply not to vowels and consonants but to the

components of vowels and consonants, called features. The idea goes back

to Roman Jakobson and is a striking, universal trait of human language. In

chapter 2 we noticed that the three pronunciations of -ed in walked, jogged, and patted are determined by the final consonant of the stem:

d if the verb ends in t or d

t if the verb ends in p, k, f, s, sh, ch, or unvoiced th

d if the verb ends in a vowel or in l, r, m, n, b, g, v, z, j, zh or voiced th

Each list is not just any old collection of consonants. The two consonants in

the first line, t and d, are pronounced with the same parts of the mouth

(tongue tip against gum ridge) and in the same way (stopping the flow of air

and then releasing it).

These features—“place = tongue-tip,” and “manner =

stopping”—are also found in the -ed suffix. A similar pattern of sharing

also is found among s, z, and the plural suffix, all of which are sibilants. What we never see is a rule such as “Add -og if the word ends in z, r, or k, ”

and other ragtag assemblies of phonemes. All this can be captured if we have rules apply to features rather than to

phonemes. One rule can state, “At the end of a word, insert i to separate adjacent consonants that have similar features for place and manner of

articulation”—no listing of t, d, s, z, -ed, or -s is necessary. Similarly, it’s obtuse to have a rule that puts a t after consonants in the list p, k, f, s, sh, ch, and th—all these consonants obstruct the air stream (they are obstruent) and

all are unvoiced. The rule can simply say, “At the end of a syllable, copy the

voicing feature from one obstruent consonant to the next.” All the unvoiced

consonants like p and k will automatically get a t, while all the voiced

consonants like b and g automatically get a d. Moreover, this rule also generates the s and z variants of the plural suffix. Simple rules that inspect, flip, or excise features are ubiquitous in the

world’s languages. They not only are more economical than rules that hack

away at arbitrary lists of consonants and vowels, but also allow speakers to

generalize. The sound ch is not normally found in English words, but any

English speaker who labors to pronounce the celebrated composer’s name

as Bach knows that if there were a verb to out-Bach, as in Handel out-Bached Bach, the past tense would be pronounced bacht, not bachd or bach

d. A rule that specifies t after “unvoiced consonants” automatically

embraces the unvoiced ch and tells a speaker what to do, even if the speaker

had never learned that ch belongs on the list. This economy and power also accrues to rules that fiddle with vowels if

the vowels are dissolved into features such as these:

Is the tongue hump at the front or the back of the mouth?

Is the tongue hump high or low in the mouth?

Are the lips rounded or not?

Is the vowel long or short?

Is the vowel tense (tongue root scrunched forward) or lax?

Take sing—sang and sit—sat. The ĭ is not replaced by some random vowel,

such as the one in say or boat or shoe; it is replaced by ă, which is identical to ĭ except for tongue height: ĭ is front, unround, short, lax, and high; ă is front, unround, short, lax, and low. A rule that simply said, “For the

following irregular verbs, lower the vowel” would have to tinker with only

one feature, not five; it would explain why ĭ was replaced by a similar

vowel, not just any old vowel; and it would work right out of the box to

yield eat—ate and choose—chose, which also lower a vowel. This simple

rule, Lowering Ablaut, is one of the three irregular rules that Halle and

Mohanan’s theory gets away with. The other two are Backing Ablaut,

which replaces mid front e in bear with mid back o in bore, and Shortening Ablaut, which replaces the long vowels in flee and shoot with their short

counterparts in fled and shot. If you have been skimming this last paragraph silently, everything

should look shipshape—the vowels in flee-fled and shoot—shot are literally

spelled as long and short versions of the same thing: ee versus e, oo versus

o. But if you have been pantomiming the sounds with your mouth and

listening to them with your mind’s ear, or better yet, pronouncing them out

loud, you should be thinking: Now wait a minute!!

As we saw in the last

chapter, “long” and “short” have been misnomers in English at least since

the Great Vowel Shift in the fifteenth century, when people scrambled the

pronunciations of vowels. The sound in flee is not a drawn-out version of

the sound in fled, nor is shoot just a lengthy shot. In pronouncing ee (ē) the tongue is higher than it is when pronouncing ĕ and more tensed up, and the

vowel glides up to a little y at the end, making it a diphthong (a succession

of two vowels pronounced as if they were one).

Likewise, in pronouncing

oo ( ū), the tongue is tenser and higher than when pronouncing ŏ, the lips are

rounded, and there is a little w sound at the end. Those vowels are not

particularly similar, and a rule capable of replacing one with another is

capable of doing almost anything. To say that a Shorten-the-Vowel rule

simplifies the hairy irregular verbs sounds like a hoax. Chomsky and Halle realized this of course, and their solution is the

most radical claim of the theory. When it comes to syntax, Chomsky is

famous for proposing that beneath every sentence in the mind of a speaker

is an invisible, inaudible deep structure, the interface to the mental lexicon. The deep structure is converted by transformational rules into a “surface

structure” that corresponds more closely to what is pronounced and heard. The rationale is that certain constructions, if they were listed in the mind as

surface structures, would have to be multiplied out in thousands of

redundant variations that would have to have been learned one by one,

whereas if the constructions were listed as deep structures, they would be

simple, few in number, and economically learned (see The Language

Instinct, chapter 4, 120–124).

Less well known is that Chomsky and Halle

made a similar proposal for the sounds of words. Each word has a deep

structure—in jargon, an underlying form—that may not sound like the way

it is pronounced; indeed, it may be unpronounceable. Phonological rules

then convert it to the surface form that is articulated and heard. In the case of the so-called long vowels in English, Chomsky and Halle

proposed that the underlying forms really are long versions of the short

vowels. That is, in the mental lexicon the vowels in the following pairs are

identical in every respect except how long it would take to pronounce them,

with the long vowels taking about twice as long:

Short Vowels

Their Long Counterparts

din

diviin ( divine)

den

sereen ( serene)

pat

saan ( sane)

fund

profuund ( profound)

shot

shoot ( shoot)

bomb

coon ( cone)

In their actual pronunciations these pairs do differ in how long it takes

to say the vowel, but they differ in many other ways besides, so why

assume that the mind lists only the difference in length?

The reason,

according to Chomsky and Halle, is that the other differences are redundant

and predictable, hence unnecessary to list. No pair of English words differs

only by vowel length. Long vowels are also tense and diphthongs (they

glide to a different vowel at the end); short vowels are lax and not

diphthongs. A theory that gave the mind the ability to store every nuance of

an English vowel—length, tenseness, tongue position, lip rounding,

diphthongs, and so on—would falsely predict that English could contain

long lax vowels, lax diphthongs, short tense vowels, and so on; and it would

allow the language to contain short and long versions of otherwise identical

vowels—all counter to fact. In a better theory, the vowel inventory would

contain only the number of vowels necessary to distinguish words (for

example, to distinguish bit from bet), and other rules would fill the vowel

out into a full set of stage directions for the mouth and throat. The best way

to do this is to list certain pairs of vowels as differing only by being long or

short, and to have the long versions trigger obligatory rules that flesh out

the rest of their pronunciation. Chomsky and Halle therefore proposed that English has a rule of Long

Vowel Tensing, which tenses all long vowels, and a rule of

Diphthongization, which adds the little y s and w s that give us the two-part

vowels in lake (leh-eek), glide (gla-eed), need (nee-y’d), loud (la-ood), and road (ro-ood).

All this is more or less unexceptionable— something in the

mind of an English speaker enforces a correlation among length, tenseness,

and being a diphthong, and the predictable details of pronunciation need not

be stored in individual lexical entries. The theory that a word is stored as an abstract, not-directly-

pronounceable deep structure has another advantage, pointed out by the

linguist Aditi Lahiri and the psychologist William Marslen-Wilson. 19

Consider the seemingly simpler idea that memory holds the actual

pronunciation of a word. The problem is: which actual pronunciation of the

word?

The word hand, for example, might be pronounced h-a-n-d when we

enunciate it carefully and distinctly, but in natural conversation it comes out

quite different. The nd is pronounced as nj in hand you, as m in hand me, and as the ng sound in hand care. (That is one of the reasons why computer

speech recognition systems, though pretty good at recognizing words in

isolation, are still poor at recognizing words in connected speech.) But if, as

Chomsky and Halle proposed, the dictionary entries for words are

schematic—so that the last segments of hand are listed, say, as “nasal” and

“dental” rather than as n and d, and they are fleshed out into full consonants

by rules that work in different ways in different contexts—then a single

representation could embrace the hand that appears in hand, hand me, hand

you, and hand care. But Chomsky and Halle went much further than merely claiming that

words have a form in memory that is not identical to their pronunciations. They proposed that the underlying form of a word can be wildly different

from its pronunciation. In particular, they proposed a complicated rule of

Vowel Shift that raises or lowers the long vowels, reenacting in the minds of

modern English speakers the Great Vowel Shift of the fifteenth century. In

other words, they claim that ontogeny recapitulates phylogeny, and that the

deep structures of words in our mental dictionaries correspond to the way

Chaucer would have pronounced them (even though Chaucer, if he traveled

through time to our century, would sound like a German to our ears).

According to the Chomsky-Halle theory, the mental representations of

words in different centuries over the past millennium, and in all the modern

dialects of English, are the same; English has changed primarily by adding

phonological rules. And English spelling, which did not track the Great

Vowel Shift or other changes in pronunciation as the dialects evolved,

captures our underlying mental representations of words. Chomsky and Halle pursue the implications. Everyone agrees that a

good spelling system ought to be stable across time and space. We should

be able to read the writings of our great-great-grandparents, and of people

on the other side of the Atlantic, even if they pronounce words differently

from the way we do. Also, a spelling system ought to encode only the

information necessary to identify the content of a word, not the trajectories

of lip pursing and tongue flicking that can be predicted from the content and

that people automatically execute as they talk. By these criteria, Chomsky

and Halle concluded, English spelling is not only exonerated of the charge

that it is an illogical, sadistic mess, but “comes remarkably close to being an

optimal orthographic system.” 20 Optimal for us, optimal for other modern

English dialects, and optimal for all the recorded dialects of the past several

centuries!

21 Hear that, all you orthographically challenged, spell-checker-

dependent, solecism-prone students and writers?

Forget cough and rough and dough and plough, and ghoti spelling fish, and George Bernard Shaw’s campaign to reform English spelling, and all the other complaints about

crazy English:

A moth is not a moth in mother,

Nor both in b other, broth in brother,

And here is not a match for there,

Nor dear and fear for bear and pear. And then there’s dose and rose and lose

Just look them up—and goose and choose,

And cork and work and card and ward,

And font and front and word and sword,

And do and go and thwart and cart—

Come, come, I’ve hardly made a start!

A dreadful language?

Man alive!

I’d mastered it when I was five. And yet to write it, the more

I tried, I hadn’t learned at fifty-five.22

What led Chomsky and Halle to this shocking conclusion?

It was the

drive to extirpate any trace of needless redundancy and complexity in their

grammar for English sound patterns. Now we really do have a clean

Shortening Ablaut rule for breed—bred, flee—fled, shoot—shot, and lose—

lost. The underlying form of breed has a double-length version of the e in

bred, so the shortening rule creates bred in a single step. (The formerly

inconvenient fact that breed itself is not double-length bred is now handled

by Vowel Shift, which makes it briid, followed by Tensing, which makes it

breed, followed by Diphthongization.) Now if it was only a handful of

irregular verbs that benefited from the Vowel Shift rule, the savings would

be paltry compared to simply stipulating that “ē changes to ĕ,” and so on. The savings begin to mount, however, when we look at other rules that can

be simplified in exactly the same way, that is, by applying to the deep, pre-

Shifted versions of vowels. With Vowel Shift available to handle the details

of the long vowel, each of the following processes can be captured as a

simple change-the-length rule:

Once you have the freedom to equip people with abstract underlying

forms for their words, the irregulars get simpler and simpler. Run—ran—

run can be handled by the rules for sing—sang—sung, drink—drank—

drunk, and so on—if you suppose that the underlying form of to run is

really to rin, and that Backing Ablaut and other rules apply to the stem, not

just the participle, to make it surface as run. Likewise, the past-tense forms

of come, give, slay, and catch are better behaved if their underlying stems

are kēm, gēv, slē, and kĕch. Most creatively of all, Chomsky, Halle, and Mohanan proposed that the

ch of Bach is a covert English phoneme that lives underground in the

lexical entries for buy and fight—namely, bēch and fēcht—and in the half-baked past-tense forms for seek and teach. Of course the chs must be assassinated before they see the light of day, but not until they have

triggered a rule that makes the past-tense form come out right. Bēch gets a -

t, changes to bôcht by the Lowering and Backing Ablaut rules, at which

point the cht triggers Cluster Shortening to yield bŏcht before ch makes the

ultimate sacrifice, resulting in the form we spell bought. What are we to make of this bold theory?

As I mentioned at the start, a

theory that posits rules for irregulars can account for the similarities

between stems and their past-tense forms, such as why swing and swung are

80 percent the same: The rule targets a vowel for change, and leaves the rest

of the verb alone. The Chomsky-Halle-Mohanan theory pushes the

performance of rules to new heights, because their rules target only certain

features of a vowel for change (such as tongue height or vowel length) and

leave the rest of the vowel alone, too. Similarly, a theory positing irregular

rules can account for the similarities in changes undergone by different

verbs, for example, why the ĭ-ă pattern in sing—sang is also found in drink

—drank and sit—sat: A few rules are shared by many verbs. Here the

Chomsky-Halle-Mohanan theory succeeds with a vengeance, forcing

almost 165 verbs to share only three rules. Any theory that can tame the quintessentially unruly English irregular

past-tense system with only three rules, each delicately adjusting a single

feature, is undeniably brilliant. But is it true?

Not necessarily. One problem

comes from the assumption that every scintilla of patterning in the verb

system needs an explanation in terms of the psychology of speakers, in

particular that the patterns are distilled out into rules in the mind. Chomsky,

Halle, and Mohanan’s rule-by-rule derivations often recapitulate the history

of a past-tense form in English over the centuries—deliberately—and that

brings to mind an alternative explanation used throughout chapter 3: that the patterns are fossils of rules that died long ago. The surviving past-tense

forms, semilawful though they are, could simply be memorized by today’s

generation without any help from the rules. The defunct-rule explanation has an advantage over the Chomsky-

Halle-Mohanan theory. Children don’t hear underlying forms, and they are

not provided with lessons about the rules that turn them into audible surface

forms. They hear only the surface forms. If the rules and underlying forms

are to play some role in mental life, children must infer the cascade of rules

that generated the surface form, run it in reverse, and extract the underlying

form. And the suggestion that English-speaking children hear run and infer

rin or hear fight and infer the German-sounding fēcht is, frankly, beyond belief. First, why would the child bother if the rules are there only to generate

the surface form, and the child already has the surface form?

(It’s different

with sentences, where the child needs rules to generate an infinite number

of new ones; with word roots, there are only a finite number to learn.) And

even if the child wanted to ferret out rules and underlying forms, how could

they ever find the right ones if the crucial clues—the ones linguists

themselves use to discover the rules—are found in pairs of words the

children will learn only in adulthood if ever, such as serene and serenity,

manager and managerial, kinesis and kinetic?

At one point Chomsky and

Halle concede the problem and say that their grammar is only what children

would construct if, hypothetically, they could hear the entire vocabulary in

one sitting before figuring out the rules, rather than learning the everyday

words first. But then it’s not clear what their theory is a theory of—it is not,

by their concession, a theory of how real children acquire words or how real

adults represent them. It may be interesting to indulge in a thought

experiment of what an optimal child ought to do if he or she had the entire

language to mull over at once. But that exercise would be useful only if the

hypothetical child were a good idealization of a real child, and that is far

from clear. Why would real children be equipped with an ability to extract

intricate chains of rules and arcane word entries if they could never put that

ability to use in the real world, and if the net result is the same language as

the one they do acquire in the real world?

It is more likely that children

store words in the mental dictionary in a form that is not radically different

in content from what they hear (though it may be more schematic).

Worse, it’s not so clear that the thought experiment would come out the

way Chomsky and Halle suppose it would. The word pairs that motivate the

strange underlying entries, such as kinesis—kinetic and intervene—

intervention, are inkhorn words encountered in writing or in the

conversations of literate professionals. Anyone who needs to use these

vowel patterns in a new word has the advantage of having seen similar

words in print. People who are literate in English have been trained, usually

with much weeping and gnashing of teeth, to associate the sounds ă and ā

with the letter a, the sounds ĕ and ē with the letter e, and so on, when they learned the alphabet. That means that when speakers have to make a choice

from among the short vowels in pronouncing new words such as in

contravene—contravention or elide—elision, they may be guided by their

knowledge of the alphabet, not by a naturally acquirable rule of English

phonology. But perhaps the biggest problem is that the Chomsky-Halle-Mohanan

theory cannot explain the third kind of similarity running through the

irregulars: the similarities among stems, as in sting, string, sling, stink, sink,

swing, and spring. In their theory the rule of Lowering Ablaut for the

participle is connected to these verbs by fiat—the entry for each verb says,

“Apply Lowering Ablaut to Me.” But then it is an unexplained coincidence

that all the verbs are so alike. The list of verbs assigned to the rule could

just as easily have been till—tull, wish-wush, fib-fub, and pith—puth. How can a theory that relentlessly soaks up every droplet of redundancy between

stems and pasts, and between the changes applying to one stem and the

changes applying to another stem, be so oblivious to the massive

redundancy among all the stems undergoing a change?

Also, how is the

speaker supposed to generalize the rules to new verbs if they are

constrained to apply only to the currently stipulated ones?

The obvious way to handle these families is to distill out their common

denominator and attach it as a condition to the rule. In the ĭ-ŭ family the ĭ

vowel tends to be preceded by a consonant cluster and followed by ng. The

consonant ng can be further analyzed into the features nasal (pronounced

through the nose) and velar (pronounced with the tongue against the soft

palate or velum).

Perhaps, then, the rule should be “Lower the vowel from ĭ

to ŭ if the stem has the pattern consonant-consonant-ĭ-velar nasal

consonant. ” Unfortunately, this rule would make errors both of commission

and omission. It would falsely include bring—brought and spring—sprang,

which do not change their vowels to ŭ, and it would falsely exclude stick—

stuck and spin—spun, which do. These verbs obviously belong in the class,

but each one violates the condition by an eyelash. The k of stick is not a nasal velar like ng, but it is a velar, pronounced at the same place in the

mouth. The n of spin also is not a nasal velar, but it is a nasal, pronounced

through the nose, just as ng is. 23

The problem, first pointed out by the linguist Joan Bybee and the

psychologist Dan Slobin, is that the irregular clusters are family

resemblance categories.24 They don’t have strict, all-or-none definitions that specify which verbs are in and which verbs are out. Instead they have

fuzzy boundaries and members that are in or out to various degrees

depending on how many properties they share with one another. String and

sling are prototypical members of the ĭ-ŭ class, packing into one word all

the consonants that are prevalent in the family. Spin and stick each misses

by a different feature; dig—dug and win—won are farther toward the

periphery; and sneak—snuck, drag—drug, skin—skun, and climb—clumb

are in a muzzy zone at the edge where speakers differ as to their

acceptability. No rule can cleanly pick out the ĭ-ŭ verbs, which is why

Chomsky, Halle, and Mohanan didn’t bother looking for the conditions that

triggered each rule but resorted to listing the verbs individually. The other irregular families work in the same way. For example, blow—

blew, grow—grew, and throw—threw are stereotypical ow—ew verbs, but

the rule for the class cannot demand that a word conform to the condition

consonant-consonant-ō.

Know—knew is in the family but misses the rule by

one consonant, draw—drew and fly—flew miss by a vowel, and slay—slew

and crow—crew are neither clearly in nor clearly out, but muzzy. Membership in an irregular family is also probabilistic when it comes to

people generalizing a pattern to new verbs. Dialectal irregular forms tend to

be close in sound to many members of a family. For example, bring—brang

is close to sing—sang, ring—rang, spring—sprang, drink—drank, and

shrink—shrank, and write—writ is close to bite—bit and light—lit. Fei Xu and I found that children’s creative irregulars work the same way. The

childhood error swing—swang is close to sing—sang and all the rest; sleep

—slep is close to feed—fed, bleed—bled, meet—met, and so on. 25 Bybee and Moder even quantified the effect by presenting their adult volunteers

with nonsense words that varied in similarity to the typical members of the

ing—ung family. Spling and skring fall smack in the middle of the family,

and about 80 percent of the participants came up with forms like splang,

splung, skrang, and skrung. Krink, trig, and pling are less similar, and only

about 50 percent of the people suggested krunk, trug, or plang. Vin, sid, and

kib share only a vowel with the verbs in the family, and only about 20

percent of the people provided forms like vun, sud, or kub.26

Chomsky, Halle, and Mohanan have tweaked rules for maximum

performance, but at a steep price. They were forced to make incredible

claims about the mental entries of words, and their theory cannot handle the

fuzzy and statistical—but psychologically active—patterns of similarity

among the verbs undergoing a rule. The irregular patterns are just not very

rulelike, and call out for something very different. When the psychologists David Rumelhart and James McClelland

announced their artificial neural network model of the past tense in 1986,

the reaction was sensational.27 Here was a model with none of the paraphernalia of linguistics—no words, no rules, no modules—but it

acquired several hundred regular and irregular past-tense forms, generalized

their patterns to new verbs, and made errors such as breaked and comed,

just like children. COMPUTERS MIMIC BRAIN IN TEST, said a headline

in the Chicago Tribune. A TURNING POINT IN LINGUISTICS, ran the

title of a review in the Times Literary Supplement.28 The implications were

“awesome,” said the reviewer, because “to continue teaching \[linguistics] in

the orthodox style would be like keeping alchemy alive.” Rumelhart and

McClelland’s model helped to launch a new school of cognitive science

known as connectionism or parallel distributed processing, which explains

mental processes in terms of networks of interconnected simple units that

vaguely resemble neurons (brain cells).

29 Many researchers saw

connectionism as a paradigm shift or scientific revolution in the study of the

mind. 30 Neural networks also became a fad in artificial intelligence and soon were put to use in picking stocks for mutual funds and controlling

expensive Japanese appliances like rice cookers and washing machines. No one doubts that language is computed by networks of neurons in the

brain. Rules—even the pristine, logic-like rules of Chomsky and Halle—are

intended as high-level descriptions of processes or structures that are

implemented in some way in neural circuitry. The difference between

connectionism and generative grammar lies in the kinds of mental

operations that are thought to be implemented in neural networks. In

particular, connectionism differs from generative grammar in the way that

associationism differs from symbol manipulation. It lacks combinatorial

rules organized into modules, and instead tries to accomplish intelligence

using Hume’s law of contiguity (if A appears with B, associate them) and

his law of resemblance (if C looks like A, let it share A’s associations).

A

neural network that works this way is called a pattern associator memory or

a perceptron. Here is how Rumelhart and McClelland’s model of the past tense

works. Despite my clash-of-the-Titans buildup, the model actually shares

some important design features with the Chomsky-Halle theory. The input

to the model is the sound of a verb stem, and the past tense is computed

from it. That is different from a model that computes a past-tense form

directly from the meaning of the verb and the concept of pastness. So

Rumelhart and McClelland are committed to at least one module—a

morphology box—sitting between meaning and sound. As with Chomsky

and Halle, a single kind of machinery is charged with computing the past-

tense forms of all verbs, regular, irregular, and suppletive ( go—went); the

verbs sit on a continuum of regularity from completely predictable to

completely arbitrary. Past-tense forms are composed piecemeal out of

miniregularities that are shared among verbs, so that sleep—slept combines

the vowel change in feed—fed and the suffixation in burn—burnt. Rumelhart and McClelland also import the standard Chomskyan

assumption that speech sounds are represented in the mind not as phonemes

but as bundles of features such as “voiced” and “nasal.”

But everything else is different. Here is the heart of the model, the

pattern associator memory:

The left-hand column is the input layer, where the verb stem is entered,

it contains 460 vaguely neuronlike units, each of which can be either on or

off. Each unit represents a tiny stretch of sound that might appear in an

English verb, such as a high vowel between two stop consonants, or a back

vowel followed by nasal consonant at the end of the word. The beginning

and end of a word are symbolized by open and close brackets (“\[” and “]”).

There are no units for individual verbs; a verb is entered by turning on the

units for the sounds it contains. As a result, similar-sounding verbs share

representational real estate. Most of the units that are turned on when shrink

is fed in are also turned on when drink is fed in (consonant cluster at the

beginning of the word, high vowel between two sonorant consonants, and

so on).

These units have no idea which word they are currently

representing. The right-hand column has an identical bank of units, and they represent

the output of the model: the sound of the past-tense form. Every input is

connected to every output by a synapselike connection that can vary in

strength, from strongly excitatory (an input signal tends to turn the unit on),

to neutral (an input signal has no effect), to strongly inhibitory (an input

signal tends to turn the unit off).

In effect each connection is a probabilistic

microrule that states something like, “If the stem contains a stop consonant

followed by a high vowel, the past-tense form is likely to contain a nasal

consonant at the end.” With 460 input units connected to 460 output units,

we have 460 x 460 = 211,600 microrules in all. When an input unit is

turned on, it sends a signal down all its lines to the output layer, where the

signal is multiplied by the strength of each connection and fed to that output

unit. Whether a given output unit turns on depends, in a probabilistic way,

on the sum of the signals that feed into it and on its own level of

triggerhappiness or threshold. The higher the summed signal is above the

threshold, the more likely the unit is to turn on; the lower the summed

signal is below the threshold, the more likely the unit is to turn off. In the neonate network the connections have strengths of zero, so the

output layer is completely off, regardless of the input. The connections then

are changed in a learning procedure, in which the model is “taught” with a

set of verbs and their correct past-tense forms. Of course, Rumelhart and

McClelland do not actually believe that a schoolmarm has to drill children

with verb conjugations. They assume that children, when hearing a past-

tense form in their parents’ speech, recognize that it is the past-tense form

of a familiar verb, dredge the verb stem out of memory, feed it into their

past tense network, and silently compare their network’s output with what

they just heard. Skeptics might wonder how a child is supposed to do all

this without the benefit of the lexical and grammatical machinery that

Rumelhart and McClelland claim to have made obsolete, but let’s put that

aside for now. The learning procedure works like this. The correct form from the

parents is displayed in a special layer of “teacher” units. The model

compares its output, unit by unit, with the correct output ( walked for walk,

came for come, and so on).

The model then adjusts the connection strengths

a tiny amount up or down depending on the difference (see the figure on the

opposite page).

If a unit is off (say, the unit for ă), and the teacher says it should be on

(because the correct past-tense form is rang), the model has to make the

input word ( ring) more likely to turn on that unit in the future. All of the

connections from incoming lines that are currently active are strengthened

an iota, and the ă unit’s threshold is lowered an iota, making it more

triggerhappy. In contrast, if a unit is on (for example, the unit for t) and the

teacher says it should be off (because the correct past-tense form is rang),

the model has to make the input word less likely to turn on that unit in the

future. All of the connections from incoming lines that are currently active

are weakened an iota (possibly driving the connection down to a negative or

inhibitory value), and the unit’s threshold is raised an iota, making it less

triggerhappy. The model is trained on a list of verbs and their past-tense forms,

presented over and over and over. A given connection will be buffeted up

and down by successive verbs in a training run, but eventually it will settle

on the strength value that does the best job, in combination with the other

connections, of producing correct past-tense forms. The network’s

knowledge of the various verbs and their past-tense forms is smeared across

the 211,600 connection strengths; one cannot point to a circumscribed part

of the network that implements a particular word, a particular irregular

family, or a regular rule. Rumelhart and McClelland trained their network on a list of 420 verbs

presented 200 times, for a total of 84,000 trials. To everyone’s surprise, the

model did quite well, computing most of the correct sound stretches for all

420 verbs. That means that a single set of connection strengths was able to

convert look to looked, seem to seemed, melt to melted, hit to hit, make to made, sing to sang, even go to went. Then Rumelhart and McClelland challenged the network with 86 new verbs, which it had not been trained

on: a test of generalization or productivity like the wug-test, the sine qua

non of rules. The model offered the correct past-tense form with -ed for

about three quarters of the new regular verbs, and made reasonable

overgeneralization errors such as catched and digged for most of the new

irregulars. Even more impressively, the model mimicked some of the tendencies of

children as they acquire English. At one point in training it produced errors

such as gived for verbs that it had previously produced correctly. It also

analogized new irregular verbs to families of similar-sounding old irregular

verbs; for example, it guessed cling—clung, sip—sept, slip—slept, bid—bid,

and kid—kid. It produced blends such as gaved and stepted that also occasionally come out of the mouths of children. It was less tempted to tack

-ed onto an irregular verb from a large family, such as feel, than onto an irregular verb from a small family, such as blow. And it was bashful about

sticking -ed onto verbs that already end in t or d, a common reluctance of

human beings that we observed in chapter 2.

Rumelhart and McClelland’s pattern associator memory is not made of

some miraculous wonder tissue. It works by one trick: Rather than

associating a word with a word, it associates the properties of a word—its

phonological features—with the properties of another word, and thereby

enjoys automatic generalization by similarity. That is, rather than

associating drink with drank, it associates dr with dr, dr with rang, ring with rang, ink with ank, and so on. At the same time, it negatively

associates dr with nked, ink with nked, and so on, inhibiting the incorrect regular form drinked. Crucially, these associations are superimposed across the different

words in the training set. When the model trained on drink is then trained

on shrink, it strengthens many of the same connections, such as ring with

rang and ink with ank. That makes shrink easier to learn—most of its connections have been prestrengthened—and it makes subsequent family

members, such as sink, easier still. It’s a short step to generalize to verbs

that have not been trained at all, such as stink—the ing—ang connections

have already been strengthened, and the ing—inged connections have

already been weakened. The same trick works for the regular verbs: When

the model is trained on walk—walked, it strengthens connections between

alk and alked, restrengthens them when trained on talk—talked, and

automatically generalizes them to stalk-stalked. The only difference

between regular and irregular verbs is that the regulars are more plentiful,

more diverse, and more consistent in the patterning of their past-tense

forms. With thousands of strong connections conspiring to turn on a t or d

at the end of a word, the model’s first tendency will be to output a regular

form. The mainspring of the model, then, is forming associations between

features and features, and that duplicates the human habit that embarrassed

the words-and-rules theory: generalizing irregular patterns to similar words. The key idea is not original to Rumelhart and McClelland. Associating

features with features is inherent to the design of many associationist

theories, and goes back to the eighteenth-century English physician and

philosopher David Hartley. 31 Hartley pointed out that if the brain

represented the properties of an object individually, then Hume’s two laws

of association—contiguity and similarity—could be pared down to one law,

contiguity. Similarity is nothing but shared properties, so associations

among properties give you generalization-by-similarity for free. That is, if

an association between bread and nourishment is in fact stored in the brain

as a set of associations between beigeness, sponginess, savoriness, and

nutrition, then when we encounter cake, which is also beige and spongy,

“nutritious” pops into mind automatically; no extra device in the brain has

to register the fact that bread and cake are similar and make a point of

transferring associations from one to the other. So does traditional linguistics, with its words and rules, have the status of

alchemy?

Not yet. In 1988 the linguist Alan Prince and I published a paper

in the journal Cognition that went after the pattern associator model

hammer and tongs. We pointed out many facts about human language that

the model, and the connectionist approach to language in general, ignored

or mishandled. 32 Other trenchant critiques appeared around that time or in

the years since.33 In a recent book the mathematician and former Scientific

American columnist A. K. Dewdney lumped connectionism with N-rays,

cold fusion, and psychoanalysis as case studies of “bad science” in need of

debunking.34 That is unfair, but connectionism has been overhyped and its

problems as a theory of the mind are real. Some of the problems might be obvious from the dissection of the

language system in chapter 2.

First, Rumelhart and McClelland’s pattern

associator memory is a device that only produces past-tense forms. You

cannot turn the arrows around and get the model to run backward and

recognize past-tense forms. Obviously people do both. Not only can we say

walked, but when we hear walked we know it means walk in the past. Children are not separately trained to produce -ed and to understand -ed. The most straightforward explanation is that they learn rules and lexical

entries, a database that can be accessed equally well by a module that sends

commands to the tongue and a module that interprets sounds coming in

from the ear. Second, the model computes every detail of the pronunciation of the

past-tense form. Yet we saw that many of these details, such as the choice

among -t, -d, and id, are found in fifteen different parts of the language system. Surely they are computed by a single phonology module that is fed

by the output of morphology and syntax, not duplicated by an amazing

coincidence in fifteen different networks, one for the past tense, one for

plurals, and so on. Third, by forgoing the use of lexical entries and relying entirely on a

word’s sound to compute its past-tense form, the model cannot tell the

difference between two words that have the same sound. It must give them

the same past-tense form, and that won’t work for soundalike verbs like

ring—rang and wring—wrung, break—broke and brake—braked, or meet—

met and mete—meted. One might reply that the problem could be fixed by

adding a few units to the input that represented the meanings of words, in

addition to their sounds. For example, a unit for “striking” could turn on

ang while a unit for “squeezing” turned on ung, differentiating ring from wring. But as we saw in chapter 2, the meanings of words don’t

systematically predict their past-tense forms: Hit, strike, and slap are similar in meaning but have different past-tense forms; take, undertake, and

take a leak are different in meaning but have the same past-tense forms. It’s

the raw fact that word 1 is not the same as word 2 which is not the same as

word 3 that triggers the different idiosyncratic past-tense forms, and that is

the distinction captured by lexical entries. Soundalike words with different

plurals and pasts are widespread in English and give rise to many of the

quirks that occupy letters to the language mavens—why a baseball player is

said to have flied out to center field, why the hockey team in Toronto is

called The Maple Leafs, why the plural of Walkman is often Walkmans. The

answer involves a beautiful design feature of human language that we will

explore in chapter 6 and that is quite unlike the knee-jerk associations that

drive the Rumelhart-McClelland model. 35

A fourth problem is that Rumelhart and McClelland had to use some

jiggery-pokery to get the model to duplicate children’s stages of language

development. We will take a closer look at how children really do learn to

use and misuse the past tense when we examine language acquisition in

chapter 7.

36

These troubles are all payback for the connectionists’ distaste for

carving a complex computational problem into a few simpler ones that can

be farmed out to mental modules optimized for each. The problems could

be solved by building separate networks for morphology, phonology, and

the lexicon, much as in traditional linguistics but with the boxes fleshed out

as neural networks. 37

But there is one problem that cannot readily be solved by dividing up

the computation into modules. It lies at the very core of the pattern

associator model, and diagnoses the main flaw in the centuries-old theory of

association-ism. The problem could not be more basic: How do you

represent an entity made of parts in a fixed arrangement, such as a word?

Units can only be on or off; you can’t inscribe them with symbols as if they

were pads of paper or bytes in a computer. The first solution that comes to

mind is to make the units into a phonetic alphabet. Assign one unit to ā, one

unit to ă, one unit to b, one unit to d, and so on. Then simply turn on the

units that spell out the word:

But this is a nonstarter. Information about the order of phonemes is lost: pit

would be indistinguishable from tip, Spiro Agnew from grow a penis. If

that’s all there were to words, you would be solving anagrams every time

you opened your mouth. A better solution is to have an array of phoneme units, one bank for the

first phoneme in a word, one for the second phoneme, and so on, up to the

longest word that a person would ever be called on to remember:

That solves the anagram problem, but it runs up against two new ones. First, how long is the longest word that the array should accommodate?

Long enough for antidisestablishmentarianism, the longest word in

standard dictionaries?

For floccinaucinihilipilification, the longest word in

the Oxford English Dictionary?

What about great-great-great-grandmother,

great-great-great-great-great-grandmother, and so on?

There is no longest

word, so something is wrong with a representation that forces us to decide

what it is a priori. The other problem is that the representation has a bank of units for the

first phoneme in a word, a bank of units for the second phoneme, and so on,

aligning the words in memory by their first phoneme, that is, left-justifying

them. But the human mind does not count off phonemes from left to right

when it perceives similarities among words and generalizes accordingly. The most tantalizing generalizations in the irregular past tense system are in

the ing—ang—ung family, with ring—rang and drink—drank and spring—

sprang reinforcing each other and inspiring fling-flang, bring—brang, and

spling—splang. But in a left-to-right array, the first three verbs do not

overlap at all:

Nothing that has been learned for ring, such as that it has a past tense with

ang, will transfer to drink or spring; the two words are no more similar than

bird and clam. But people must find them similar, for the irregular system is

rife with families of words that don’t align properly at their left edges, such

as dive, drive, strive, and most obviously, prefixed forms such as stand,

withstand, and understand or come, become, and overcome. The same problem fouls up any generalization that depends on the ends of words, and

we know there is a huge one: the choice of -t, -d, or id, which hinges on whether the last phoneme is voiced, unvoiced, or a t or d. The last phoneme

in a left-to-right representation can be position 2 (for the verb add), position

3 (for ask), position 4 (for risk), and so on, all the way up to position 23 for

floccinaucinihilipilify and beyond. A left-to-right representation would have

to learn how to pronounce the suffix separately for every length of word. Rumelhart and McClelland must have recognized this problem, because

they came up with a creative alternative: The units stand for things they

called Wickelphones, named after the psychologist Wayne Wickelgren who

first conceived them.38 A Wickelphone is a sequence of three phonemes,

like ipt or str. English has about forty phonemes, and if we add the special symbols “\[” and “]” for the beginning and end of a word, there are about

67,000 possible Wick-elphones, each needing a unit. By representing a

word by its Wickelphones, one sidesteps both the anagram problem and the

left-alignment problem. For example, strip contains the Wickelphones ip],

rip, str, tri, and \[st. You don’t have to worry about their order, because they

snap together in only one way: \[st at the beginning, then str, then tri, and so

on. And the Wickelphones for rip overlap the Wickelphones for strip, so

their representations are similar, just as the human mind perceives them. (Rumelhart and McClelland in fact wanted to represent words in terms of

their Chomsky-Halle-esque features, rather than their phonemes, so a unit

actually stood for three features in a row, such as stop-high-stop or voiced-

unvoiced-voiced—a Wickelfeature.)

It may seem hard to believe that the simple act of registering a word in

an unstructured bank of units is a near-insoluble problem, but it is:

Wickelphones, though ingenious, don’t work either. The human mind cares

about single phonemes and the features that compose them, and the

Wickelphone has submerged them into unbreakable chains of three-in-a-

row. For example, silt and slit have no Wickelphones in common. The first

dissolves into \[si, sil, ilt, and lt], the second into \[sl, sli, lit, and it].

But people clearly hear them as similar, as we see in historical changes in

English such as brid becoming bird and thrid becoming third. Worse, in some languages the Wickelphone cannot represent certain words at all. The

Australian aboriginal language Oykangand has a word algal meaning

“straight” and a word algalgal meaning “ramrod straight,” and they are

made up of identical Wickelphones: alg, al], gal, lga, and \[al:

Since units are either on or off, they have no way of representing two of

something, and the Wickelphone theory therefore incorrectly predicts that

Oykangand speakers should not exist (nor the speakers of many other

languages; words like these are not uncommon).

39

Also, a theory of how the mind represents things should predict what

the mind finds easy and what it finds hard. The easy tasks should be

computed by simple operations on the representation, the hard tasks by

lengthy sequences of operations. Here, too, the Wickelphone makes the

wrong prediction. When linguists explain to their classes how human

languages use some kinds of rules and not others, they almost always use

the same example: that no language has a rule that flips a word to its mirror-

image, say, forming the plural by converting tip to pit, gum to mug, and dog to god. But a Wickelphone-to-Wickelphone network can do exactly that,

and quite easily: strengthen every connection between a Wickelphone ABC

in the input and its mirror-image Wickelphone CBA in the output, and

weaken all the other connections. Not only is mirror-reversal easy, but it is no harder to learn than the

simplest conceivable relation between input and output: copying the stem

verbatim, which involves strengthening the connections between ABC and

ABC. The only difference between mirror-image reversal and verbatim

copying is that we, the theorists peering into the model, can read the unit

labels and see that ABC goes to CBA in one case and to ABC in the other. But the model cannot read its own node labels; all it cares about is the

consistency of the input-output relations, and they are the same in both

cases. Likewise, all kinds of crazy rules, such as replacing all as with b s, all

bs with cs, all cs with ds, and so on, are as easy to learn as copying the input

to the output. This is not just a quibble; it explains an embarrassing lapse in the

performance of Rumelhart and McClelland’s model. The model was mute

when asked for the past tenses of simple but somewhat unusual-sounding

words, like jump, pump, warm, and trail. And it garbled several others,

turning squat into squakt, tour into toureder, and mail into membled. The lapses are puzzling to us because intuitively nothing could be simpler than

copying a stem over to the past-tense form before adding -ed. But a pattern

associator memory has no placeholder called “stem” that can be copied, and

no operation to do the copying. All it does is associate sounds with sounds,

and if the training set happens to be missing words with certain

combinations of sounds such as -ump or -ail, the model will be at a loss,

and will sit in silence or cough up a hairball of bits and pieces that are

vaguely associated with the sounds it has been trained on. 40

All the problems go away if you bring back the rationalist theory that

the mind manipulates symbols organized into hierarchical structures by

rules. A verb such as to outstrip might be represented something like this:

The phonemes are held in their correct order by a treelike scaffolding that

embodies the morphological structure of the word (how it is built out of

stems, prefixes, and suffixes) and the phonological structure of its parts

(how they are built out of chunks like onsets, rimes, vowel nuclei,

consonants and vowels, and ultimately, features).

The similarity to other

words such as strip, restrip, trip, rip, and tip falls mechanically out of the

fact that they have identical subtrees, such as an identical “stem” or an

identical “rime.” And computing the regular past-tense form is nothing but

attaching a suffix next to the symbol “verb stem”:

It doesn’t matter whether the underbrush dangling beneath the “verb stem”

symbol is walk, outstrip, jump, pump, or bftsplk—if you have a mental

symbol “verb stem” and know how to put a suffix next to it, the entire

vocabulary of verb stems lies waiting at your feet. Finally, since a tree

structure is built out of recursive rules (for example, “a stem can combine

with a prefix to form a new stem”), no length limit needs to be set

beforehand, and words of any length such as re-outstrip or great-great-

great-grandmother can be represented. Symbolic trees require fancier neural hardware than the smooth purée of

units that are popular among connectionists, but those models hardly do

justice to the brain anyway. Recently, a few neural network modelers have

shown how hierarchical trees can be implemented in more organized neural

networks. 41 One conjecture is that the periodic rhythms of neural firing, long downplayed in neuroscience, serve as the glue that binds together the

units that represent an abstract slot in a tree and the units that represent its

content. For example, when the units for the “coda” slot fire at twenty times

a second, and the units for p fire in synchrony with them, also at twenty

times a second, the system as a whole knows that the coda is p. Simultaneously, the units for “nucleus” can be firing thirty times a second,

and the units for i can be firing in synchrony thirty times a second, and the

system knows that the nucleus is i. The units for “coda,” “nucleus,” p, and i

are all active at the same time, but the system doesn’t get confused and

think that i is in the coda, because the shared firing patterns link each sound

to its slot. That theory may or may not be right, but I mention it to show that

abstract symbols and complex structure are not incompatible with plausible

neural network models. After Alan Prince and I took apart the pattern associator model, the

linguists breathed a sigh of relief because they thought they didn’t have to

learn neural network modeling after all, and the connectionists dropped it

like a hot potato. So it is ironic that Prince and I are probably the model’s

biggest fans. It does, after all, explain a major phenomenon that rule

theories ignore, and it accounts for not one but several aspects of children’s

language development. In comparison, the twenty-five connectionist

models of the past tense that have been devised in reply to our critique have

been disappointments, not one of them anywhere near as ambitious as the

original.42 Many sidestep the Wickel-phone problem by using a Dick-and-

Jane version of English that contains only monosyllabic words made of a

consonant, a vowel, and a consonant, such as walk and run. Others

implicitly concede that words are composed of symbols for stems and

symbols for affixes and don’t even bother computing a past-tense form. They merely select from an innate menu of five or six units that stand for

the five or six suffixes or vowel changes in the language. Some other

mechanism then has to apply the suffix or vowel change to the stem to get

an actual past-tense form. That unmentioned mechanism, of course, is what

we call a rule. Many modelers beef up the network with an intervening

layer of units hidden between the input and the output layers, but direct

benchmark tests find little or no improvement.43 Each of the inventors has

added a different patch that narrowly fixes some problem that Prince and I

pointed out—what programmers call a hack or a kluge—but none defends

his brainchild as an actual theory of how that part of the mind works. And

no one has made an empirical prediction or accounted for several kinds of

data in the way that Rumelhart and McClelland did. One phenomenon, two models, both explaining too much to be completely

wrong, both too flawed to be completely right. Prince and I have proposed a

hybrid in which Chomsky and Halle are basically right about regular

inflection and Rumelhart and McClelland are basically right about irregular

inflection. Our proposal is simply the traditional words-and-rules theory

with a twist. Regular verbs are computed by a rule that combines a symbol

for a verb stem with a symbol for the suffix. Irregular verbs are pairs of

words retrieved from the mental dictionary, a part of memory. Here is the

twist: Memory is not a list of unrelated slots, like RAM in a computer, but

is associative, a bit like the Rumelhart-McClelland pattern associator

memory. Not only are words linked to words, but bits of words are linked to

bits of words. The bits are not Wickelphones, of course, but substructures

like stems, onsets, rimes, vowels, consonants, and features, perhaps

something like this:

Furthermore, the nodes of one word (such as string) overlap the same

nodes in other words (such as sling, stick, stink, and swim).

As a result, irregular verbs show the kinds of associative effects found in a

connectionist pattern associator. People find families of similar irregular

verbs easier to store and recall because these verbs repeatedly strengthen

their shared associations. And people occasionally generalize the irregular

patterns to new, similar verbs, because the new verbs contain material that

already had been associated with the pattern from the old verbs. Prince and I were not the first to modify the words-and-rules theory in

this way. Many generative linguists have been uncomfortable with the

Chomsky-Halle ethos of using industrial-strength rules to account for

everything that is systematic in language. Mark Aronoff, Joan Bresnan, Ray

Jackendoff, Rochelle Lieber, Andrew Spencer, and others have suggested

that language uses two kinds of rules: true rules that speakers generalize

freely, and lexical redundancy rides that merely capture patterns of

similarity among words stored in memory.44 A memory system in which

patterns of similarity are registered and occasionally generalized is simply a

pattern associator memory, and Rumelhart and McClelland have given us a

sketch of how one might work. The modified words-and-rules theory may sound like a sappy attempt to

get everyone to make nice and play together, but it makes a strong

prediction. The prediction is that regular and irregular inflection are

psychologically, and ultimately, neurologically distinguishable. But how

could they be distinguished if both involve patterns that people can

generalize?

The answer is that irregular inflection depends on memorized words or forms similar to them, but regular inflection can apply to any

word, regardless of whether the word is readily retrievable from memory. Regular inflection has that power because it is computed by a mental

operation that does not need access to the contents of memory, namely, a

symbol-processing operation or rule, which applies to any instance of the

symbol “verb.” The evidence will be woven through the rest of the book as

we explore how words are used in conversation and in reading, how new

words are created, how children learn their mother tongue, how language is

organized in the brain, and whether the languages of the world conform to a

universal design. We will see how in dozens of cases of language use that

have nothing in common except a failure of access to memory, irregular

patterns are disabled but the regular rule works fine. If the modified words-and-rules theory is correct, it would have a

pleasing implication for the centuries-old debate between associationism

and rationalism: Both theories are right, but they are right about different

parts of the mind. 5

WORD NERDS

When we use our native language, a torrent of words flows into and

out of the brain. The occasional frustration of having a word stuck

on the tip of the tongue, the slow ordeal of composing a passage in a

foreign language, and the agony of a stroke victim struggling to answer a

question remind us that our ordinary fluency with language is a precious

gift. This chapter looks at how words and rules pop into mind as we use

language in real time. The topic offers a good test of the modified words-

and-rules theory introduced at the end of chapter 4.

According to the theory,

regular forms are generated by rules, and irregular forms are retrieved from

memory; the memory, however, is not a list of slots but is partly associative,

linking patterns with patterns as well as words with words. The theory differs from Chomsky and Halle’s theory of generative

phonology, with its battery of rules that generate both regular and irregular

forms. It also differs from Rumelhart and McClelland’s theory of

connectionism, with its pattern associator memory that stores both regular

and irregular forms. We saw how each model has problems with the facts of

the English language. Chomsky and Halle, in leaching every bit of

patterning out of memory and concentrating it in rules, had to propose

implausible deep structures for words, and could not explain why irregular

verbs come in families of similar forms. Rumelhart and McClelland, in

dismantling every bit of structure in language, had to propose clumsy

Wickelphones to represent words, and could not explain why regular verbs

are copied so reliably into their past-tense forms. In this chapter we will turn from the mechanics of these classic models

to the general principles that power them. After all, some of their problems

might come from details of implementation, which could always be

improved in later models. Here we will see whether the intellectual core of

each model—that memory is compressed to a minimum, in the case of

Chomsky and Halle, and that generalization works by the laws of

association, in the case of Rumelhart and McClelland—fits the facts of

human language use. The great challenge for any theory of language is productivity, the

ability to generate and understand an unlimited number of new forms. One-

piece words like duck and walk can always be listed in memory, but they

are not the only words we trade in; we are constantly faced with new words

formed by combining prefixes, stems, and suffixes. In a language such as

Kivunjo or Turkish every word may come in a half a million to several

million forms, and speakers could not possibly have memorized them all in

childhood. Even in a morphologically challenged language like English we

have to cope with new word forms every day. The psychologists Harald

Baayen and Antoinette Renouf calculated that every time you open a

newspaper you will be faced with at least one word with un- that you have

never seen before, one with - ness, and one with -ly: words like uncorkable,

uncheesy, headmistressly, breathcatchingly, pinkness, and outdoorsiness.1

And these are just three of the forty-odd common prefixes and suffixes in

English. Coping with new word forms is a problem for all language users, not

just human language users. Computer programs that understand and

produce conversational English need morphology modules to deal with

novel forms of words.2 Even my lowly spell-checker boasts one, according

to its documentation:

spell collects words from the named files, and looks them up in a

hashed spelling list. Words that do not appear in the list, or cannot

be derived from those that do appear by applying certain inflections,

prefixes or suffixes, are displayed on the standard output. The morphology module allows it to store just the stems of words and to

compute the inflected and derived forms by rule. I asked the program to

check the first two paragraphs in this chapter and report on its activity. About one out of every seven words was missing from its dictionary and

had to be computed with rules:

Similarly, the program can handle psycholinguistic jargon such as

overregularized and underlyingly without complaining:

overregularized = over + regular + ize + d

underlyingly = under + lying + ly

If I add mosh and Bork to its dictionary, I don’t have to add moshed, moshes, moshing, Borked, Borks, and Borking as well; the spell program

passes the wug-test. How does the human mind handle new inflected and derived words?

Does it work like my spell-checker, with a minimal dictionary and a set of

rules for carving unfamiliar words into prefixes, stems, and suffixes?

Does

it rely on an enormous dictionary that lists all the common forms of every

word?

According to the updated words-and-rules theory, the mind has rules

for regular forms and relies on a pattern-associating memory for the

irregular forms. The evidence for this hybrid model, we shall see, is that

when people use an irregular form, they must have that form or similar

forms in memory, whereas when they use a regular form, they don’t need to

access memory at all. A simple property of memory is that the more often you hear something the

better you remember it. Uncommon words, therefore, have weak memory

entries and should be harder to retrieve. The words-and-rules theory

predicts that rarity should hurt an irregular verb, but not a regular verb. The

first place to look for this effect is in the statistics of the English language

itself. Here is a Top Ten list, the ten most common verbs in English:

Verb

Number of occurrences in a million

words of text

1.

be

39,175

2.

have

12,458

3.

do

4,367

4.

say

2,765

5.

make

2,312

6.

go

1,844

7.

take

1,575

8.

come

1,561

9.

see

1,513

10.

get

1,486

The verbs are ranked by their frequencies in a million-word corpus of text

assembled from newspapers, magazines, popular books, textbooks, and

other sources, and analyzed at Brown University by the computational

linguists Nelson Francis and Henry Kuèera. 3 You may have noticed

something the verbs have in common: They are all irregular. Indeed, some

are really irregular. The top four are the only verbs that are irregular not just

in the past tense but also in the present tense: be-is/are, have-has, do—does,

and say—says. And the #1 and #6 spot contain verbs that are so irregular

that their past-tense forms are different words altogether: be—was/were and

go—went. There cannot be a Bottom Ten list for the least common verbs in

English, since the truly rare ones won’t turn up even in a million-word

corpus. But we can look at the rarest verbs that do turn up, the 877 verbs

tied for last place with a frequency of one in a million. Here are the first ten

in alphabetical order:

Verb

Number of occurrences in a million

words of text

abate

1

abbreviate

1

abhor

1

ablate

1

abridge

1

abrogate

1

acclimatize

1

acculturate

1

admix

1

adulterate

1

As you can see, they are all regular. To be exact, 860 of the one-in-a-million

verbs, 98 percent, are regular. Another sixteen are prefixed irregulars that

are parasites on their far more common roots: bethink, forswear, inbreed,

misread, outdraw, outfight, overbear, overdrive, overlie, overwrite, presell,

regrind, spellbind, unbend, unbind, unwind. Only one word among the

rarities is an irregular root: smite. Irregular verbs are the most common verbs and vice-versa, in English

and in most other languages. 4 The explanation is simple. Irregular forms have to be memorized repeatedly, generation after generation, to survive in

a language, and the commonly heard forms are the easiest to memorize. If

an irregular verb slips in popularity, a generation of children will fail to hear

its past tense often enough to remember it. Since the rule “Add -ed” can

apply to an item regardless of its frequency, the children will use the regular

suffix, and that verb will be regular for them and for all subsequent

generations. Those who cannot remember the past are condemned to

compute it. Joan Bybee did some historical digging to prove this conjecture. Remember that Old English had about three times as many strong irregular

verbs as Modern English, including obsolete forms such as cleave—clove,

crow—crew, abide—abode, chide—chid, and geld-gelt. Bybee looked at

thirty-three strong verbs that survived in Modern English, and divided them

into verbs that remained irregular and verbs that became regular. She then

looked up their modern frequencies in the Brown corpus. The still-irregular

verbs appear an average of 515 times per million words; the regular

defectors appear an average of 21 times.5

We can feel this force of history acting today when we look at the past-

tense forms that are low in frequency. Smite is the only one-in-a-million

verb root clinging to an irregular form, smote. But no one can use it in

conversation with a straight face; the form is sliding out of the language

before our eyes. Likewise, heave-hove, stave—stove, rend-rent, bid—bade,

slay—slew, smell—smelt, and thrive—throve—thriven are a bit peculiar, and

one can predict they will go the way of chid and crew. Sometimes a form is

familiar enough to block a regular version, but not quite familiar enough to

sound natural, and speakers are left without any good past-tense form for it. Complete this sequence: I stride, I strode, I have \_\_\_\_\_\_.

Most people

grimace at this point, equally uncomfortable with stridden and strided. Stridden may be found in dictionaries and occasionally in prose (as in,

“where...

a pinnacle of beauty had stridden the earth,” from Rebecca

Goldstein’s 1989 novel The Late-Summer Passion of a Woman of Mind),

but it is not to be found in the million words of the Brown corpus. It hovers

in the mists of memory, tainting strided without stepping onto the stage

itself. Crucially, this never happens with regular verbs. Complete this

sequence: I agglutinate, I \_\_\_\_\_\_, I have \_\_\_\_\_\_.

Agglutinate, like smite,

appears once in a million words, and not in the past tense. Here, though,

people have no reluctance to supply the elusive past and participle forms

agglutinated. The other pastless verbs in the sample are just as easy to

conjugate— allure—allured, badger–badgered, carouse—caroused—as are

the countless verbs that are too rare to show up even in a million words: to

fleech, to fleer, to stint, to prescind, to anastomose, and so on. Does this difference really come from the faintness of a rare irregular

past-tense form in memory, or does it come from a squeamishness about

using the rare verb in any form?

We may not use smote very often, but then

we don’t use smite every day, either. Are there any verbs that are fine in

their bare, stem form and odd only in the past tense?

That would be a good

test of the theory that verbs and their past-tense forms are listed separately

in memory: If so, one entry could be solid while the other is frail. The place

to look is in idioms, clichés, and collocations, which often are used solely in

the infinitive or the present tense. Take the verb forgo. Though uncommon, it retains a certain liveliness,

particularly in the sarcastic phrase to forgo the pleasure of, as in You’ll

excuse me if I forgo the pleasure of watching the video of your wife giving

birth. Now try to put it in the past tense. The word is certainly not forgoed,

but the alternative is not quite right either: Last night I forwent the pleasure

of watching Hank’s vacation slides. 6 Similarly, it is perfectly natural to say I don’t know how she can bear the guy, but something is odd about I don’t

know how she bore the guy. Though one might say I dig The Doors, man!

,

it’s much harder to say In the ’60s, your mother and I dug The Doors, son. And while everyone knows what That dress really becomes her means, the

past-tense version is almost unintelligible: But her old dress became her

even more. Here is another example:

ZITS reprinted with special permission of King Features Syndicate. Irregular stems and their past-tense forms can part company and accrue

different degrees of familiarity—just what we would expect if they are

separate entries in memory. The slippage can go the other way too, with an

irregular past-tense becoming more familiar than the verb itself. Recall

from chapter 3 that people often are foggy about what verb stem goes with

smitten, rent, shod, hove, and wrought. When that happens in the history of

a language, a past-tense form can lose its moorings and drift over to some

other verb. For example went originally went with wend, and now goes with

go. None of this happens to regular verbs. Some are used primarily in

negations and hence appear in the infinitive, such as He doesn’t suffer fools

gladly. But when the cliché is coaxed into the past tense, nothing interesting happens: None of them ever suffered fools gladly. Other verbs that are

common in general but rare in the past tense, such as afford and cope, also

do just fine when plunked in the past tense, as in I don’t know how he

afforded them and It’s a miracle how she coped with him. In the Z its

cartoon, the joke would disappear if the irregular hite “be bad” were

replaced with its regular synonym suck. All this is what we would expect if

people react to uncommon regular past-tense forms not by searching for

them in memory but by analyzing them into a stem and a suffix by rule. The

past-tense form would inherit the familiarity of the verb as a whole, because

the past-tense form simply is the verb (plus an ornament) as far as the mind

is concerned. We have seen that the distribution of verbs from the high end of the

frequency range to the low end tells us something about the psychology of

the people using the verbs. Equally informative are the clues found among

the verbs with the lowest frequencies of all, those that appear exactly once. There is a lovely technical term for a word that appears once in a body of

text: a hapax legomenon, plural hapax legomena, Greek for “once said.”

The term comes from philology, the study of old texts. Hapax legomena can be a nuisance to scholars of ancient languages

because with only one instance of a word one can never be sure what it

means. But Harald Baayen has shown that they can be a gold mine for

linguists interested in whether a prefix or suffix is truly regular and

productive. Baayen devised a formula to capture in a single number the

productivity of a suffix (or any other form): the number of hapax legomena,

that is, the number of words with that suffix that appear exactly once in a

corpus, divided by the number of times the suffix appears in the corpus

summing over all words. 7 Here is an intuitive way to understand it. Suppose you want to know how many fish are in a huge body of water. Obviously you can’t count them all. You get your rod and reel and pull out

a fish, tag it, release it, and give it time to swim away. Then you catch

another fish, tag it and release it, then another, and so on, noting for each

fish whether you had caught it before. If the body of water contains a small

number of fish, at some point you’ll keep catching the same few fish again

and again. If it has an enormous number, most fish you pull out will not

have been caught before. If the number is open-ended, or essentially infinite

—if the fish breed faster than you can catch them—you will never catch the

same fish twice. After a million tries, you can write down the number of

fish that you have caught once, the number you have caught twice, and so

on. The larger the proportion of the catch that consists of fish caught only

once, you may conclude, the more fish are out there in the water. The body of water is the English language. A fish is a suffixed word. A

million casts of the fishing rod is a million-word corpus. A fish that has

been caught ten times is a word with a frequency of ten per million. A fish

that has been caught only once is a word with a frequency of one: a hapax

legomenon. If the creel is filled with hapax legomena, the words must be

breeding quickly. That is, the suffix must be productive and the set of words

accepting it open-ended. The vague notion that a rule of language is

“productive” or “open-ended” can therefore be translated into a number. What happens when we cast our line into English and pull out regular

past-tense forms?

Of the 15,369 catches in the Brown corpus, 871 are hapax

legomena, regular past-tense forms pulled out only once. That means that if

one were to keep fishing (adding to the Brown corpus), 5.7 percent of the

new regular past-tense forms would never have been seen in the past tense

before. Is that a high rate or a low rate?

The best comparison is to a class of

words we know to be relatively unproductive, namely, verb stems such as

eat and walk. New ones such as mung and mosh are created not by a rule but by occasional moments of inspiration in creative wordsmiths. When we

bait our hooks for verb stems, we get 170,931 catches, of which 877 are

hapax legomena, a rate of new word formation of only 0.5 percent. (Actually, this is an overestimate, because it includes new words formed by

derivational prefixes and suffixes such as uncorkable and pinkness. ) That is

a tenth of the value for the regular suffix, and shows that we breed new

regular past-tense forms far more quickly than we invent verbs. The other

interesting fishing expedition is for irregular past-tense forms. We land only

62 hapax legomena among 10,832 catches, a rate of 0.6 percent—

essentially the same as the rate for verb stems. 8 The statistics confirm that

regular forms are generated freely, presumably by a rule, whereas irregular

past-tense forms are stored in the same manner as ordinary words. So far I have been doing psychology in a roundabout way—by working

backward from the statistics of English vocabulary, now and then asking for

your gut reactions to verbs in various parts of the frequency range. But the

same effects can be shown in systematic studies. Michael Ullman and I

asked ninety-nine people to rate their gut reactions to several hundred verbs

and their past-tense forms, each presented in a sentence, on a scale of 1

(unnatural) to 7 (natural).

9 We got their ratings of the bare verb stems, as

well as their ratings of the past-tense forms, so that we could disentangle

any queasiness about a past-tense form—say, maimed or smote—from

queasiness about the verb itself, such as to maim or to smite. Ullman found that ratings of irregular past-tense forms depended on

their frequency in the language: The more common the verb form, the better

people liked it. (This was true even when he controlled for the familiarity of

the verb itself, using a standard statistical procedure.) Ratings of regular

past-tense forms, in contrast, did not depend on their frequency in the

language—a rare verb form such as maimed was just as natural sounding as

a common one like walked (again, controlling for the fact that maim is less

natural than walk).

Instead, ratings of regular past-tense forms depended

highly on the ratings of the stems: The more natural the stem, the more

natural the past-tense form. That is just what we would expect if people saw

the past-tense form as simply the stem itself, with a decoration added by a

rule. People’s ratings of the irregular past-tense forms correlated less well

with their ratings of the stems, which is what we would expect if irregular

past-tense forms are separate words, which are only linked to their stems. The same patterns hold when people have to cough up past-tense forms

under time pressure, as they do in rapid conversation. Sandeep Prasada,

William Snyder, and I brought another group of volunteers into the lab,

seated them in front of a computer screen that flashed verb stems at them,

and asked them to blurt out the past-tense form as quickly as they could. 10

A microphone connected to a voice-operated trigger sent a signal to the

computer. By timing the interval between stimulus and response, the

computer could estimate how long it took people to read the verb, mentally

compute the past-tense form, and begin to say it aloud. We chose verbs in

pairs, one used frequently in the past tense and one used less frequently, but

both used equally often in the nonpast forms. (As before, we wanted to hold

constant the familiarity of the verb, and compare only the past-tense forms.)

For example, the stems ring and strive are both used about seven times per

million words of text, but rang is used twenty-one times whereas strove is

used only four times. Similar statistics hold for the regular items pour and

soak, which are used equally often in the stem form, but poured is about

twenty times as common as soaked among past tense forms. All the verbs

were presented in random order. With irregular verbs, the more frequent past-tense forms came out of the

people’s mouths faster, as we would expect if they were stronger in memory

and therefore quicker to retrieve. But with regular verbs, the more frequent

past-tense forms were no faster than the less frequent ones, which suggests

that people were not retrieving them preformed from memory but were

assembling them on the fly. The effects were small—only a few hundredths

of a second—but we found them in three experiments, and four other teams

of experimenters have replicated the results. 11

When you produce an irregular form, you not only have to dredge it out

of memory but also must repress the “Add -ed” rule so you don’t say

breaked or broked. Linguists call this principle blocking—the irregular form

blocks the rule—and the experiments help us understand how the mind

implements it. One possibility is that when we need to utter a past-tense

form we first scan our list of irregular verbs to see if it is there, and if it

isn’t, we turn on the rule. That predicts that the slowest irregular verb (the

one at the end of the list) should be faster than the fastest regular verb. The

prediction is wrong: Irregular forms usually are slower to produce than

regular forms; they are never faster. 12

A more likely possibility is that words and rules are accessed in parallel,

that is, at the same time. As we plan to utter a verb in the past tense, we

simultaneously look up the word in memory and activate the rule. An

inhibitory link runs from the memory box to the rule box, which gradually

slows down the rule as evidence for a match is found, and eventually turns

it off. Memory lookup is not an alphabetical search, of course, like finding a

word in a real dictionary. The phonemes and syllables in a word contact

their counterparts in memory piecemeal, more and more of them finding a

match as the milliseconds tick by. As soon as all the pieces match some

entry, the irregular form linked to the entry is fetched and shunted to the

vocal tract. While the lookup is in progress, the inhibitory signal sent to the

rule box gets stronger and stronger, and when all goes well, the rule is

braked to a halt. Occasionally the matching or fetching goes awry and the

rule runs to completion, producing a speech error, such as I carefully looked

at ’em and choosed that one. 13

If a memory entry is faint or blurry because the word is uncommon, the

matching and fetching will be especially erratic, and often the rule may not

be braked in time. For example, many people may produce slayed as a

speech error (it should be slew), and their listeners may tuck it away in

memory as a genuine past-tense form for the verb. If that happens often

enough, doublets such as slew/slayed and strove/strived will come into

common parlance. Eventually the irregular form may disappear outright,

like chid, dempt, and abode. Ull-man found that verbs with doublet past-

tense forms ( strived and strove, dreamed and dreamt, dived and dove) have lower frequencies than verbs with just one past-tense form. Ullman also got

snapshots of the weakening grip of these irregular forms and the rising

strength of their regular counterparts: Among the doublets, the lower the

frequency of the irregular form, the better-sounding the regular form. 14

The flip side of failing to recall an irregular form is tripping a false

alarm for one when the verb is in fact regular. Word lookup is not

instantaneous, and as it proceeds a few irregular verbs in memory might

crudely match a regular probe. That could temporarily slow down the rule

until the last jots and tittles of the word are properly matched and the false

matches have petered out; only then will the rule be allowed to proceed

unhindered. This predicts that regular verbs that are similar to irregulars,

inviting temporary false matches, should be slower to produce in the past

tense. The psychologists Mark Seidenberg and Maggie Bruck found exactly

that: Regular verbs such as smell, greet, and bake (which rhyme with

irregular tell, meet, and make) are slower for people to utter in the past tense

than regular verbs such as walk and look, which don’t rhyme with any

irregular and therefore have an unimpeded run through the rule box.15

Incidentally, there is no contradiction between saying that regular past-tense

forms don’t depend on their memory entries and that they can be slowed

down by temporary false matches with other verbs’ memory entries. From

your brain’s point of view, no verb is either regular or irregular until it has

been looked up in memory and discovered to have, or to lack, a special

past-tense form. The mental machinery for words also can be studied in the laboratory by

asking people to recognize words, as they would when listening or reading. The problem with studying word recognition is that it happens privately

inside the skull, and it’s not clear at what point in this process people may

be said to have “recognized” the word. Is it when they know they have seen

the word before?

When they know what it means?

When they know how to

use it in a sentence?

The experimental psycholinguists who call themselves

“word nerds” usually try to tap the moment at which people are willing to

say that the word is a word and not just something that looks like a word. Volunteers see or hear a mixed-up series of real words and fake words, such

as narse or bluck, and have to press one button if the item is a word and another button if it is not. This task, called lexical decision, doesn’t

correspond to anything that people do outside the lab, and no one really

knows what goes on in people’s heads as they do it, but hundreds of

experiments have used the procedure.16 The task is popular because it tells

us something about how the mental dictionary is organized. If people are given a word and then given it again, they are faster at

recognizing it—that is, discriminating it from a nonword—the second time. Apparently a mental dictionary entry can be primed (sensitized or prepared)

to match an instance of a word; the effect is called repetition printing. Interestingly, a word also can be primed, though not as strongly and not for

as long, by a word that is associated or similar in meaning to it, such as

doctor for nurse or duck for goose. Words must be hot-linked in memory, like pages on the World Wide Web, so that when one word is turned on, it

becomes easier to turn on related words. What happens if the priming word is an inflected form and the target

word (the one the person has to recognize) is its stem?

That is, what

happens when walked is presented as the priming word and then walk is

presented as the target?

According to the words-and-rules theory, if the

prime is a regular form, the mind should analyze it as a stem plus a suffix,

and the stem should prime the mental dictionary entry just as if the stem

had been presented by itself. That is, walked should have the same effect on

recognizing walk as walk itself does, because as far as the mind is

concerned, the word walked just is the word walk with a suffix. In contrast,

an irregular form like swept should be seen as a word that is distinct from

its stem sweep, though hot-linked to it. Therefore, while swept might prime

sweep, it should do so less effectively than sweep itself, because as far as

the mind is concerned, swept is its own word, not sweep with a vowel

change and a -t. The psychologist Robert Stanners and his colleagues found

exactly that, and four other laboratories have replicated the finding,

including one that bypassed the button-push and measured the brain’s

response to the words directly via electrodes pasted to the scalp. (We will

return to this technique in chapter 9.)17

You may be objecting that walked and walk overlap in letters more than

swept and sweep do, and perhaps it’s just the brute repetition of black lines

in the shape w-a-l-k or a repetition of the sounds of wŏk that primes the

word walked, and not anything so airy as a putative entry in a mental

dictionary. But these studies include controls such as market as a prime for

mark or gravy as a prime for grave—words that overlap as much as regular

forms and their stems do but are not related in meaning or grammar. Nothing much happens; the mind is not impressed by mere overlap in

letters or sounds. Another way to show that priming occurs deep in the mind and not in

the senses is to play a tape of the priming word spoken aloud, and then flash

a printed word on a computer screen. If a noise coming in from the ears

helps a person to recognize a squiggle coming in from the eyes, what must

have come between them was a representation of the word in the mind:

The psychologists William Marslen-Wilson and Lorraine Tyler did the

experiment. 18 They found that when people heard asked, it was easier to read ask, but when they heard gave, it was no easier to read give. In a diabolical twist, the researchers ran the experiment again using a technique

called subliminal priming. The priming word is flashed very quickly on the

screen and is immediately covered up by a row of random squiggles. Subjects insist that they can’t see a thing. But the word must have registered

unconsciously in their brains, because invisible filled made them faster at

recognizing fill. In contrast, invisible rode did not make them any faster at

recognizing ride. As usual, mere overlap— fillet and fill, rude and ride—had no effect. Is an irregular form, then, simply related to its stem the way duck is

related to goose—as two words that just happen to be similar in meaning?

Apparently not. A stem and its irregular past, even if they are stored

separately, have to be treated as members of the conjugation of a single

word; otherwise there is no reason that rode would block the generation of

rided. Marslen-Wilson and Tyler in fact showed that irregular forms are

tightly linked to their stems, not just loosely related like duck and goose. A

word can prime a word with a similar meaning only if the second one

appears immediately after the first; the priming by shared meaning dies

down quickly. A word can prime itself, however, over many seconds,

minutes, sometimes even hours or days. Marslen-Wilson and Tyler tried a

variant of the priming experiment—a spoken word followed by another

spoken word, rather than by a word flashed on a screen—and found that

regular and irregular forms were equally effective in priming their stems. (No one knows why irregular forms can prime their stems in the allsound

version of the task, but not in the sound-then-screen version.) Words with

similar meanings, such as gold and silver, primed each other as well. But

grammar proved to be a stronger tie than meaning when the target did not

follow the prime immediately but only after a delay of 35 seconds. By then,

gold was a distant memory and an ineffective prime when silver came

around, but gave continued to prime give, and of course filled continued to

prime fill. That suggests that the mind represents gave not as a word that just happens to be similar in meaning to give but as a separate morpheme

that represents the verb give in the past tense.19

These differences between regular and irregular verbs were nicely predicted

by the words-and-rules theory, but do they rule out the alternatives?

Mark Seidenberg has suggested that one piece of evidence for rule use

—the finding that uncommon regular past tense verbs are no slower to

produce than common ones—is in fact a signature of connectionist pattern

associator memories.20 Pattern associators are designed to ignore words

and remember patterns. Regular verbs are so plentiful that an uncommon

one like stalk is bound to share patterns with many other regulars, such as

stop and walk, so its rarity should not hurt it. Irregulars are few in number,

so if one is uncommon, it will have no similar irregular forms to lean on,

and it will be harder. Seidenberg and the computer scientist Kim Daugherty devised a

simulation much like the Rumelhart-McClelland model, but with an extra

layer of units hidden between the input and output, a modification that can

make a pattern associator more powerful:

An immediate problem in comparing this model to people is that pattern

associators compute everything in a constant number of steps, in this case

two, and therefore take the same amount of time for all verbs, regular and

irregular, common and uncommon—unlike people. So Daugherty and

Seidenberg interpreted the error score for the pattern associator (the

discrepancy between its guess and the correct answer) as an indicator of a

verb’s difficulty, and assumed that some black box fed by the associator

would take longer to assemble the “harder” verbs. Unfortunately, though

the designers can see the model’s error score, the model itself cannot. Once

the training sequence is over, the model has no way of knowing whether its

guesses are right or wrong, unless it memorizes all the words in the

teaching sequence to use as a crib sheet. But the whole point of a pattern

associator is to do away with entries for words!

In any event, the model did not act like people: To the modelers’

disappointment, it found uncommon verbs harder than common verbs,

whether they were irregular or regular. (Recall that people find the

uncommon regular verbs no harder than the common ones.) The problem

was that the irregular verbs, such as in stick-stuck, were interfering with the

regular patterns that uncommon verbs such as stalk depend on. So

Daugherty and Seidenberg retrained the model with only 24 irregular verbs

and 309 regulars, giving the uncommon regulars plenty of similar neighbors

to lean on and keeping the pesky irregulars in check. That did the trick—the

model now found the uncommon regular verbs to be not much harder than

common ones, mimicking the human being. This decimation of the irregular verbs is questionable. Adults know

about 165 irregulars, preschool children about 80, 21 and as we shall see in

chapter 8, the ratio of regular to irregular words varies wildly from

language to language and should not be precariously balanced on a narrow

ledge of values just to get the model to behave properly with English. There

is a more serious problem, however: The model’s success hinges on its

artificiality. The model is an idiot savant tailored to do one task: generate

the sound of a past-tense form. It can be made insensitive to the frequency

of regular words because it is insensitive to words, period; all it knows is

the sounds of words. Unfamiliar stalk gets dissolved into st- and -alked, and both sounds can be made familiar if there are enough regulars in the

training set and not too many irregulars. But real human beings do not live

by sound alone; they need to know what a word means. The human head

must contain something that differentiates stalk from stop and walk. And

that something—a lexical entry, or some link between the sound and

meaning of stalk—is bumped up and down in strength as the person

encounters a word more or less often. If a pattern associator were to model

people’s actual knowledge of words, not just their sounds, it too would be

accumulating information about the word as a whole, including its past

tense associations. That would make it unlikely to mimic the human pattern

of generating uncommon regular past-tense forms as easily as common

ones. What about the opposite theory, Chomsky and Halle’s all-rule model

from generative phonology?

According to their theory, regular verbs are not

tagged for the -ed rule (they undergo it by default), whereas irregular verbs

are tagged for their vowel-change rules. If the tags are strengthened by

practice, then the uncommon irregulars should be slower and less certain,

which squares with the facts. But another prediction of the theory does not square with the facts. The

theory’s main precept is that any trace of redundant information is pried out

of memory and computed by rule; the theory assumes that in the human

brain, memory is expensive but computation is cheap. The ultimate

redundancy is in the regular verbs, where storage of the past-tense forms

would be unconscionably wasteful. If predictable forms are never stored,

then surely regular verbs are not stored. In contrast, the words-and-rules theory assumes that memory is

constantly working alongside rules—that’s how irregular verbs arise to

begin with—and it would be a strange mental block indeed that would force

the memory system to be amnesic for all the regular past-tense forms it

hears. (After all, we remember them just fine in quotations such as All men

are created equal and The quick brown fox jumped over the lazy dog. ) The

words-and-rules theory predicts only that people don’t depend on stored

past-tense forms, not that they are incapable of storing them. People use a

rule to generate and judge past-tense forms when they need to, and if some

regular forms have been stored in memory, they are available but not

indispensable (indeed, the stored versions may even get in the way).22

Ullman’s experiment uncovered two cases in which people do store and

retrieve regular forms. In words that have two acceptable past-tense forms

such as the doublets dived/dove and dreamed/dreamt, people’s judgments

about the regular versions, such as dived and dreamed, depended heavily on

how often those forms are used in the language. This makes perfect sense,

even within the Chomsky-Halle theory. Suppose the gatekeeper to memory

follows the principle, “Store anything that is unpredictable.” Irregular

forms, of course, are unpredictable, so they (or their tags to the vowel-

change rules) are always stored. But the regular member of a doublet is

unpredictable, too. If you already know dreamt, the blocking principle rules

out dreamed, just as came rules out comed. Suppose now that you hear someone say dreamed. If your language processing could be displayed as a

readout on your forehead, it would say, “Dreamt plus the blocking principle

says that dreamed shouldn’t be in the language. But someone just said

dreamed, so it must be in the language after all. Since my rule system won’t

generate it, I had better remember it.”

A similar explanation works for another case in which frequency makes

a difference: regular verbs that rhyme with irregulars, such as blink (which

rhymes with drink and stink) and glide (which rhymes with ride and stride).

Human memory always holds out a temptation to generalize irregular

patterns. So when a person contemplates the past tense of blink and glide, a

little voice in the head will be whispering blunk!

and glode!

When the

person hears someone say blinked or glided, she makes a mental note of the form, because its existence comes as a bit of a surprise. 23

So the generative phonology theory could handle the data by allowing

less predictable regular forms like dreamed and blinked to be marked for

the regular rule. What about the other regular forms—ordinary past tenses

and plurals that don’t resemble irregular forms and are therefore 100

percent predictable?

The answer is that some of them appear to be stored in

memory, too. Four teams of psychologists have asked people to decide

whether a printed string of letters is a word or not, the “lexical decision”

task described earlier.24 All the studies found that people take a few

hundredths of a second longer to recognize rare regular forms than common

ones (holding constant the frequencies of the verbs or nouns themselves).

That suggests that people do build up a memory trace for common regular

forms, even though logically they don’t have to. Human memory is not a

scarce resource reserved for the incompressible nuggets that cannot be

generated by rules. If a word form is common enough, we can look it up

directly, rather than breaking it into parts and looking up the parts. Why do people use memorized regular forms in these experiments but

not in the experiments on rating and producing past-tense forms described

earlier?

Harald Baayen and the psychologist Robert Schreuder propose that

the rule system and the word system process a word form in parallel (at the

same time), as in the diagram on page 130, and that the two systems race

against each other to produce or analyze a form.25 Since each system will

be faster with some words and slower with others, people will rely more on

one system or the other in different tasks to optimize their ability to speak

or understand quickly and without error. Whether past-tense forms will be

looked up or generated by rule depends on the nature of the forms and on

what the person needs to do with them. With irregular words there is no

choice; only the word system can handle them. But with regular words, it

depends on the task and on the word. The tasks fall along a continuum. At one extreme is the leisurely

process of rating or reflecting on the naturalness of forms, or choosing the

most natural one. Since people need only determine whether a verb could

have a well-formed regular past tense in the language, and the rule is

always there to provide one, they don’t care whether they have seen the

past-tense form often, seldom, or not at all. In the middle of the continuum

is the task of producing a form under time pressure. If many of the verbs are

regular it will be fastest to let the rule generate the forms, in which case the

frequency of the stored regular past will make no difference. But if the list

contains many irregulars, speakers are being forced to go to memory so

often that they might as well use any regular form they come across, and the

fainter ones will take longer to retrieve than the stronger ones. 26 At the other end of the continuum is the task of deciding whether a string of letters

is a word. Here people must say “no” to forms that well might be words but

that they have never seen before, such as refeamed, and “yes” to forms that

they indeed have seen before. The task discourages the use of rules and

encourages matches against memory, and thus is a sensitive assay for any

trace of a word lingering in memory. That is why in this task, common

regular forms are faster to recognize than uncommon ones. It depends on the words as well. A regular form stored in memory is

useful only if it is strong enough to be retrieved quickly. Otherwise, the rule

system will strip off the suffix and retrieve the stem (especially if the stem

is common) before the inflected form can be retrieved, and the rule system

will win the race.27 The winner depends on other factors, too, such as how

easy the suffix is to strip from the stem and whether the suffix is ambiguous

because the language has a lot of syncretism (suffixes found in many

conjugational slots).

Baayen and Schreuder have formalized the parallel-race theory as a

mathematical model that predicts how long people should take to recognize

stems, plurals, and past-tense forms with different frequencies in a

language. The predictions are an excellent match to the behavior of real

humans across a large set of experiments. When Baayen and Schreuder

removed the rule pathway, making it a memory-only theory, or removed the

word pathway, making it a rules-only theory, the model’s resemblance to

humans plummeted. 28 The two routes of the words-and-rules theory appear

to be just the right number to account for actual human performance. Frequent pairing is one of the engines of associationism, and similarity is the other. Hume enshrined it as his second law of association, and

behaviorism depended on it in a version called stimulus generalization. Train a pigeon to peck at a green key, and it will peck furiously when the

key stays green, vigorously when the key is yellowish green or bluish

green, languorously when it is yellow or blue, and barely at all when it is

red or violet. The pattern is called a generalization gradient, after the shape

of the curve that emerges on a graph when the pecking rate is plotted

against the wavelength of the key. Connectionism, too, runs on similarity. Here is how Rumelhart, McClelland, and their collaborator Geoffrey Hinton

once explained the appeal of pattern associators: “People are good at

generalizing newly acquired knowledge. If, for example, you learn that

chimpanzees like onions you will probably raise your estimate of the

probability that gorillas like onions. In a network that uses distributed

representations, this kind of generalization is automatic.”29

According to the words-and-rules theory, similarity is indeed important

when people generalize an irregular pattern to a new verb, and for exactly

the reason connectionists invoke: Similar forms overlap in memory, and the

associations of one automatically transfer to the others. But similarity is not

important when people generalize a regular pattern to a new verb, because

that can be done by grammatical combination, which simply joins a suffix

to any stem labeled as a verb. Here is a way of visualizing the difference. Imagine that words fill a

space with many dimensions, each corresponding to the presence of some

sound, such as a consonant cluster at the beginning of a word or a vowel at

the end. Irregular verbs fall into neighborhoods of similar forms that carve

out regions of this space: sing—sang, ring-rang, and drink-drank cluster in

one region; hide-hid, slide—slid, and bite—bit cluster in a second; blow-

blew, know—knew, and grow—grew cluster in a third. When a verb falls in a

neighborhood or in the halo surrounding it, people are tempted to generalize

the local irregular pattern to it, as in bring—brang, fight—fit, and snow—

snew. Here is a simplified picture, in which each dot is a word, clusters of

dots are words with similar sounds, and the shaded areas are the sounds that

tempt people to generalize an irregular pattern:

Now, how do regular words fit into this space of sounds?

Do they fall into

neighborhoods of their own—neighborhoods that just happen to be bigger,

more populous, and more evenly sprinkled with words than the irregular

neighborhoods?

That is the connectionists’ explanation of why the regular pattern is more

easily generalized. But if regular forms are generated by a rule that merely

states, “Join a suffix to a verb,” the sounds of the verb should make no

difference: The rule is triggered by a symbol, “verb,” not by the sounds of

particular verbs. The potential regular territory should not be an archipelago

of neighborhoods in the territories unoccupied by irregulars. They should

suffuse the entire space of sounds, with every potential word sound as

eligible for the regular suffix as every other one. Graphically, the result

would not look like a big set of clusters but like a uniform blanket:

The words-and-rules theory predicts that a verb with any sound at all can

get the regular suffix, not only those in neighborhoods of familiar regular

words. As with the effects of frequency, the English language itself provides us

with clues. First, regular verbs aren’t intimidated by a gang of irregulars

occupying a neighborhood, but tramp right through and pitch their tents. Every irregular family in English tolerates regular interlopers:30

Irregular Family

Regular Interloper

hit—hit, split—split

pit—pitted

cut—cut, shut—shut

jut—jutted

bleed—bled,

feed—

need—needed

fed

bend—bent,

send—

mend—mended

sent

sleep—slept, keep—

seep—seeped

kept

sell—sold, tell—told

spell—spelled

bind—bound, find—

mind—minded

found

grow—grew, blow—

glow—glowed

blew

take—took, shake—

fake—faked

shook

stink—stunk, slink—

blink—blinked

slunk

The most blatant sign of regular imperialism is the ability to occupy the

same patch of territory as an irregular verb—that is, regular verbs that are

homophonous (identical in sound) with irregular families:

Irregular Verb

Regular Homophone

The dentures fit. He fitted the dentures. She rang the bell. They ringed the city. He hung the painting. He hanged the rustler. We met the passenger. We meted out the

punishment. They lay on the

They lied about what

couch. happened

The regular carpetbaggers are not just a handful of tough verbs that have

clung to their territories against steep odds; they can be created at the drop

of a hat. Here is a do-it-yourself guide for creating a regular verb with any

sound you want. English has a rule that creates a verb by prefixing out- to a

person’s name, as in It out-herods Herod (from Hamlet) or Clinton is trying

to out-Kennedy Kennedy. Make your favorite sound into someone’s name. Presto!

A regular verb. In 1983 everyone heard about Sally Ride, America’s

first woman in space. But a few years later Mae Jemison got even more

publicity for being the first African American woman in space—she out-

Sally-Rided Sally Ride (not out-Sally-Rode Sally Ride).

For many years the

most infamous prison in New York was Sing Sing. But in 1971 a riot at the

Attica Correctional Facility propelled it into the limelight: Attica out-Sing-

Singed Sing Sing (not out-Sing-Sang Sing Sing, out-Sang-Sang Sing Sing,

out-Sing-Sung Sing Sing, or out Sung-Sung Sing Sing).

There is a popular

board game in Japan called “Go.” Suppose that the Japanese continue to

emulate American culture and that Monopoly becomes even more popular. It would have out-Go’d Go (not out-Gone Go).

In the next chapter we will see experiments that catch people in the act of creating these forms. 31

Finally, we find regular verbs boldly going where no verb has gone

before, into the nooks and crannies of phonological space that rarely or

never see verbs. Few if any English verbs end in -ach, -ev, or -a, but we can

say Handel out-Bached Bach; Yeltsin out-Gorhachev’d Gorbachev; and We

rhumba’d and cha-cha’d all night. Nothing in English sounds anything like

the word oink, but since it found its way into print as a verb in 1969, no one

has had any compunctions about saying that a pig oinked. 32

Michael Ullman has documented the wanderlust of the regular suffix in

his study in which people judged the naturalness of hundreds of verbs and

their past-tense forms. 33 He found that irregular forms that belonged to large families of similar forms ( ring—rang, sing—sang, drink—drank, and

so on) were judged as more natural than irregular forms with small or

nonexistent families, such as stand-stood, holding all else constant. But

with regular forms, it made no difference: verbs from large families, like

walk, talk, stalk, and balk, were no more acceptable than verbs from tiny

ones, such as scour. 34

Can the act of slapping regular suffixes on strange-sounding verbs be

reproduced in the lab?

Sandeep Prasada and I replicated the experiment by

Joan Bybee and Carol Moder that asked people to guess the past-tense

forms of novel verbs like spling. 35 Recall that Bybee and Moder got a pigeonlike generalization gradient: The more similar a novel word was to

existing irregular verbs, the more likely the subjects were to generalize an

irregular pattern to it. For example, the pseudo-verb spling is similar to

many real irregular verbs such as sling, slink, string, and shrink, and

subjects inflected it as splang or splung 80 percent of the time. Krink is less similar, and it came back as krank or krunk only about 50 percent of the time. Vin, which is barely similar, came back as van or vun only about 10

percent of the time. Prasada and I tested nonsense verbs like those, but we also tested

nonsense verbs that varied in their similarity to regular verbs. A verb like

plip sounds like many English regular verbs, such as chip, clip, dip, drip,

flip, grip, nip, pip, quip, rip, ship, sip, slip, skip, snip, strip, tip, trip, whip,

and zip. Verbs lik e glinth and smaig don’t rhyme with any English verb roots. And verbs like ploamph and smeerg are very much unlike English

verbs because they violate the sound pattern of the English language. English phonology doesn’t allow a long vowel to precede a consonant

cluster at the end of a syllable unless all the consonants are produced by the

tip of the tongue. For example, toast is a possible English word, but toask

and toasp are not. 36

We gave these six kinds of verbs—three varying in similarity to

irregular families, three varying in similarity to regular families—to

university students and to the trained Rumelhart-McClelland pattern

associator model. For the irregular verbs the model gave a reasonably good

impersonation of a human being, showing a generalization gradient in

which only the verbs that sounded a lot like irregular verbs were readily

given irregular forms. But for the regular verbs the model and the human

diverged. People put -ed on strange-sounding ploamph at virtually the same

rate as they put it on familiar-sounding plip. The pattern associator also had

little trouble with plip, but with ploamph it could muster a regular form

only about 10 percent of the time. Sometimes it offered hopeful guesses

loosely based on irregular forms:

greem—grame

proke—prokt

brilth—brilt

Sometimes it offered a form that belonged to some other verb it had been

trained on:

brilth—prevailed

ploag—pleaded

proke—trusted

krilg—brewed

And sometimes it concocted bizarre blends:

slace—fraced

smeeb—imin

ploanth—bro

smairf—spurice

trilb—treelilt

smeej—leefloag

frilg-freezled

This sorry performance confirmed our diagnosis that the model’s errors in

Rumelhart and McClelland’s own tests— mail—membled, tour—tourder,

and silence for jump, pump, warm, and glare—were not random noise. They were a symptom of the model’s inability to generalize its training to

words that did not sound like the words it had been trained on. The failure is instructive. Pattern associator memories, unlike symbol

crunchers, cannot exploit the basic gadget of computation called a variable. A variable such as “verb” can stand for an entire class of items, regardless

of their phonological content. That allows a rule to copy over the material

of a stem and simply hang a suffix on it, whatever it is. A pattern associator,

in contrast, has to be painstakingly trained with items bearing every input

feature in the class. If a new item bearing a novel combination of features is

presented, the model cannot automatically copy over the combination; it

activates bits and pieces that are vaguely associated with the features and

coughs them up in a hairball. Connectionists were quick to blame these problems on the 1960s-era

technology used in the original Rumelhart-McClelland model. Forget the

Wickelphones, they said; we never meant for them to be taken seriously. 37

And forget the direct stimulus-response connections; the state of the art in

connectionist models has three layers of units (as in the Daugherty-

Seidenberg model described earlier), not two. The new layer, hidden

between input and output, has been proven to allow a pattern associator to

solve problems that a two-layer model cannot. For example, it can decide

whether a number is odd or even, and can apply the formula “A or B but not

both,” neither of which the old models could do. In principle, this deluxe

kind of model can be trained out of its slavery to similarity and create new

families that don’t necessarily reflect the shared features of the input. The computational linguists Richard Sproat and Dana Egedi put these

upgrades to the test. 38 They replaced the Wickelphones with input and

output layers that any linguist would be proud of, and added a sophisticated

decoder that turned the output layer into a sensible string of vowels and

consonants. They also retrofitted the model with a hidden layer of units, and

trained it using a state-of-the-art learning procedure. Like the original

model, and like a human being, it generalized irregular patterns to verbs

that were similar to the irregulars it had been trained on. But also like the

original model, and unlike a human being, it generalized poorly to novel

regular verbs. Some were left unchanged. Some were confused with other

verbs in the training set, such as train—trailed, spoke—smoked, and glow—

glanced. And fully a quarter were bizarre distortions, such as conflict—

conflafted, wink—wok, yield—rilt, satisfy—sedderded, and quiver—quess. A pattern associator’s ineptitude with novel combinations appears to be

deeply rooted in its design, not just a failing of a first-generation

implementation. Many connectionists have gone back to the drawing board,

but none has been able to get a pattern associator memory to generate new

regular forms properly. Several modelers, stymied by the models’ habit of

outputting gibberish, have hardwired various patches into their model that

are tailor-made for regular verbs. One team of modelers included a second

pathway of connections that linked every input unit to its twin in the output,

implementing by brute force the copying operation of a rule.39 Another

team added an innate clean-up network in which the units for -ed strengthen

the units for an unchanged stem vowel and inhibit the units for a changed

vowel, shamelessly wiring in the English past-tense rule. 40 And as

mentioned in chapter 4, many connectionist modelers have given up on

trying to generate past-tense forms altogether. Their output layer contains

exactly one unit for every past tense suffix or vowel change, turning

inflection into a multiple-choice test among a few innate possibilities.41 To

turn the choice into an actual past-tense form, some other mechanism,

hidden in the wings, would have to copy over the stem, find the pattern

corresponding to the chosen unit, and apply the pattern to the stem. That

mechanism, of course, is called a rule, just what connectionists claim to be

doing without. When it comes to generalizing regular inflection to novel words, pattern

associators are simply the wrong tool for the job. The problem is that a

single mechanism is being asked to do several jobs with contradictory

demands. To discriminate among similar irregular sounds with different

outputs, such as drink—drank, slink—slunk, think—thought, and blink—

blinked, a pattern associator has to cultivate an ear for the tiniest nuances of

the sound of the input and confine its associations to them. That is exactly

the opposite of what it has to do to generalize the regular pattern to novel

words, where it must be oblivious to differences in their sounds and plaster

them all with the suffix. Moreover, with the same pathway it uses to

discriminate and generalize among verbs, a pattern associator also must

record all the sounds of the stem and try to copy them to the past-tense

form. In a standard grammar these demands are handled by three parts that

stay out of each other’s way: a lexical entry (which marks a verb as

irregular), a category symbol such as “verb” (which can be joined to a

suffix by a rule), and a phonological representation (which comes through

in the output untouched).

The pattern associator has to do the jobs of all

three, none satisfactorily. Perhaps the most important lesson of the chapter is that the mind, like

any complex device, is a system of mechanisms optimized for different

jobs. Any theory that has one mechanism doing all the work is proposing a

kind of crippleware that the human brain is bound to outperform. If the

mechanism is a set of rules, it loses the advantage of cacheing the results of

frequently performed computations so that it can look them up quickly

rather than recomputing them every time. If the mechanism is a set of

associations, it loses the advantage of variables and the rules that combine

them. As the psychologist William James wrote, “Thought is...

a kind of

algebra...

in which, though a particular quantity be marked by each letter,...

it is not requisite that in every step each letter suggest to your thoughts

that particular quantity it was appointed to stand for.” 42

6

OF MICE AND MEN

Irregularity in language, the quintessence of illogic and caprice, often

inspires bouts of idle curiosity. The Boston Globe columnist John

Powers wonders:

Why do artists show Adam with a belly-button?

Why is Germany

the Fatherland but Russia the Motherland?

Who made “impact” a

word?

Why does Queen Elizabeth wear that kerchief?

How can

there be a Miss Universe pageant without Miss Pluto?

Why is Peg

short for Margaret?

If tin whistles are made of tin, what are foghorns

made of?

Why do dilemmas have horns?

If “mice” is the plural of

“mouse,” why isn’t “hice” the plural of “house”?

1

Sometimes the only answer to such questions is, That’s just the way it

is. I have no idea why in Old English mus had the plural mys but bus had

the plural husas, to say nothing of the Queen and her kerchief. But

sometimes idle questions do have answers. In the comic strip “Funky

Winkerbean,” a Little Leaguer uses his time on the bench to ponder the

mysteries of baseball (see the following page), but we can do much better

than his teammate at resolving them. The primary meaning of plate is “a smooth, flat, relatively thin, rigid

body of uniform thickness,” and according to the Rules of Major League

Baseball, “Home base shall be marked by a five sided slab of whitened

rubber fixed in the ground level with the ground surface,” which fits the

definition nicely. In the early days of baseball every pitch led to a fair strike,

a foul strike, or a ball. A foul strike was any attempt to strike at the ball that

was not fair (in bounds).

FUNKY WINKERBEAN by Tom Baliuk

The term fair strike fell into disuse, and foul strike was shortened to strike, with foul reserved for striking the ball out of bounds. Baseball managers,

unlike the coaches of other sports, sometimes have to run onto the playing

surface, to confer with the pitcher or kick dirt on the umpire’s shoes. As for

why a batter is said to have flied out—no mere mortal has ever flown out to

center field—there is an even more satisfying answer, and it is the topic of

this chapter. To begin with, yes, it really is flied out. In The Careful Writer, Theodore

Bernstein notes,

You won’t find it in most dictionaries, but flied is the past tense of

fly in one specialized field: baseball. You could not say of the batter

who hoisted a can of corn to the center fielder that he “flew out”;

you must say he “flied out.” 2

Since we are discussing baseball slang, I can’t resist explaining the lovely

phrase a can of corn, an old term for a high, lazy fly ball. It goes back to the

early decades of the twentieth century, when grocery stores stacked canned

goods on a high shelf and the grocer would retrieve a can by tipping it with

a pole or grabber and catching it as it fell, much as an outfielder catches a

fly ball. The language columnist Jan Freeman adds, “The fact that it’s corn

in the can...

is probably based on euphony— can of corn, with its neat

near-rhyme, sounds a lot snappier than can of peaches or can of beans.” 3

But back to flying out. Flied out is one of many irregular forms that

mysteriously turn up in regular garb when used in certain ways. For

example, one might say All my daughter’s friends are lowlifes, rather than

All my daughter’s friends are lowlives, even though the usual plural of life is

irregular lives. One might say I’m sick of dealing with all the Mickey

Mouses in this administration, not the Mickey Mice. Toronto has a hockey

team called the Maple Leafs, not the Maple Leaves, and when their goon

tries to decapitate an opposing sniper and is sent to the penalty box for

high-sticking, we say that he high-sticked his opponent, not that he high-

stuck him. Flying out and other systematic regularizations, we shall see, offer an

elegant corroboration of the theory that language in general, and the

regular-irregular contrast in particular, may be explained as an interaction

between words and rules. In the preceding chapter we saw how people are happy to use a rule like

“Add -ed” whenever their memory does not supply an inflected form. That

can happen for many reasons. With new coinages such as to wug, to fax, to

Bork, and to mosh there is no past-tense form in memory. With rare verbs

such as to allure, to badger, and to carouse the form may be too faint to

retrieve reliably. With strange-sounding words such as to ploamph and to

frilg there is no similar form in memory either, preventing the use of

analogy. With words with irregular homophones such as mete and lie, or

words with irregular neighbors such as wink and blink, there may be

competing forms in memory. But in all these failures of memory, people are

not left speechless; their rule can step into the breach and generate a regular

past-tense form. In this chapter we examine cases where memory is useless not for

quantitative reasons (as when a word is relatively rare or strange) but for

qualitative reasons. The forms that surprise us, such as flied out and

lowlifes, either violate the standard format of a word stored in memory or

skirt the mechanism that funnels information from memory to the rules that

compute the word’s form. The regular suffix rises to the occasion, just as it

does when a word is rare or strange. That underscores the power of a rule: It

can apply whenever memory fails, regardless of the reason for the failure. These examples also add to the debate on whether rule processing or

memory associations are the main motor of productivity in language. In the

preceding chapter we saw that the chief rival to rules, the pattern associator

memory, has trouble generating regular forms for novel and unusual words. Yet the connectionists who defend these memory models have not

conceded. The behavior of pattern associators and other artificial neural

networks depends on dozens of settings, such as the number of hidden

layers, how many units are in each one, and the nature of the training set. Tweaking these networks has become part of the neural network modeler’s

art, and any failure of a model is taken as a challenge to squeeze out more

performance by souping it up or by adjusting the richness or leanness of the

mixture of regular and irregular forms in the input. I suspect that ultimately little will come of this trial and error, because

the problem with pattern associators lies in their very design. In the next

two chapters we will see that no magical combination of settings is likely to

work for all inflections in all languages. But in this chapter, I set aside

questions about numbers. Rules will show their worth, and pattern

associators their limitations, because of the kind of information captured in

words, not how many words of various types there are. Systematic regularization immediately proves that sound alone cannot be

the input to the device that computes inflected forms, as it is in most pattern

associator memories. A given sound such as fly can come out the other end

of the device as flew and flown when referring to birds, but as flied when

referring to ballplayers. The question is: What is that extra input, and why

does it make a difference?

Many language mavens, psychologists, and connectionists have come

up with the same explanation, which can be called the semantic stretch

theory.4 It comes from the intuition that language is a direct conversion from meaning to sound, and states that the extra input features are semantic. When a verb is given an extended or metaphorical meaning, the new sense

is felt to be dissimilar from the original, and this inhibits the speaker from

using the original’s irregular form. People sense that flying out is different

in meaning from flying, and that a lowlife is different in meaning from a life, so they are inhibited from borrowing the irregular forms for those words. Perhaps they house a pattern associator augmented with units for bits of

meaning (a unit for “wing-flapping,” a unit for “sleaziness,” and so on), and

the pattern of a new word with a stretched meaning does not overlap

enough with the pattern of the word with the original meaning to parasitize

its associations. Or perhaps people are trying to make themselves clear and

worry that the irregular form will give their listeners the wrong idea, such

as that a ballplayer has acquired superhuman powers. The main problem with the semantic stretch theory is that its basic

premise is wrong: Semantic stretching in itself has no effect on a word’s

past tense or plural. There are hundreds, perhaps thousands, of examples in

which the meaning of an irregular word is stretched, sometimes to the

breaking point, and people do not abandon its irregular forms:5

• If a new word is formed from an old irregular word by prefixing, the

new word stays irregular. When eat begets overeat, the new past tense is

overate, not overeated. Similarly, we get overshot (not overshooted), undid, preshrank, remade, outsold, and so on. • New nouns constantly are being formed by compounding a word onto

an existing noun. When the input is irregular, so is the output: bogeymen,

not bogeymans; superwomen, not superwomans; also muskoxen,

stepchildren, milk-teeth. • Using a noun as a metaphor also does nothing to its irregularity. Misguided reviewers of my books attack straw men, not straw mans, and

we speak of chessmen, snowmen, sawteeth, children of a lesser god, and

being six feet under. The petroleum industry refers to freeloaders who tap

into wells and pipelines as oil mice. 6 A species of wasp that hunts for bees is called a beewolf; several of them are known as beewolves. My computer,

alluding to a spawned program called a “child process,” recently spat out

the eloquent message, sendmail\[95]: NOQUEUE: SYSERR: getrequests:

accept: No children. • English has hundreds of idioms based on irregular verbs, as we saw

in chapter 2, and they steadfastly retain the irregularity of the originals, no

matter how strained or opaque the metaphor: cut a deal, not cutted; took a

leak, bought the farm, caught a cold, hit the fan, blew him off, lost his

marbles, put him down, came off well, went bananas, threw up. These irregular loyalists also falsify the suggestion that people

regularize words to avoid ambiguity and make themselves clear.7 Many of

these idioms are ambiguous between literal and idiomatic senses, such as

bought the farm and threw up, and some are ambiguous with other idioms

as well: blew away, for example, could mean “wafted,” “impressed,” or

“assassinated”; put him down could mean “lower,” “insult,” or “euthanize.”

But that doesn’t tempt anyone to single out one of the meanings in each set

by saying buyed the farm, throwed up, blowed him away, or putted him

down. Conversely, the past tense of to grandstand is grandstanded, not grandstood, but grandstood would be perfectly unambiguous if anyone said

it. The same is true of Mickey Mice, high-stuck, and lowlives, which would

be perfectly clear, especially in context. But with these unambiguous words

people are tempted, even compelled, to use a regular past-tense form.8

It’s not that meaning is irrelevant to the abandonment of irregular forms;

it is relevant, but only sometimes and in certain ways, and meaning is not

the only thing relevant. A theory that does predict when a word will lose its

irregularity has been developed by the linguists Paul Kiparsky, Edwin

Williams, Rochelle Lieber, and Elizabeth Selkirk, with amendments from

me and my collaborators.9 It comes right out of the words-and-rules theory,

once the words part and the rules part have been fleshed out in more detail

than I have given you so far. To preview:

• Words are stored in the mental dictionary not as haphazard

bundles of information but in a standard format called a root. • Rules don’t just throw words or parts of words together; they

provide a scheme in which the properties of the new

combination can be computed from the properties of the

parts and the way they are arranged. A combination that

obeys this scheme is said to have a head. We shall see that a word that conforms to these standards—a word with a

root and a head—is well-behaved: If it looks like it should be irregular, it is

irregular. A word that violates the standards has to give up its irregular

form; these are the words, such as flied out and lowlifes, that arouse the curiosity of language-lovers. This explanation may be called the word

structure theory, because it says that the structure of a word—in particular,

whether it has a root and a head—determines whether it gets to keep an

irregular form. This chapter will show how the word structure theory explains dozens

of puzzles about English words that fill the language columns and cartoon

pages. Since it appeals to the essence of words (the root) and to the essence

of rules (the head), the success of the theory will stand as a confirmation of

the words-and-rules theory more generally. We begin with words that

cannot find their roots; then we will turn to words that have lost their heads. People are infinitely creative with the sounds they use in conversation. They salt their speech with gestures, sound effects, foreignisms, names, and

quotations, all used as if they were actual words:

So he starts to argue with me, and I just went \[rolls eyes].

When I hit the rock, the tire made a pffffffffffffft sound. This townhouse has that je ne sais quoi. I’ve been Norman Mailer’d, Maxwell Taylor’d. I’ve been Rolling

Stoned and Beatled till I’m blind.10

If he “Yes, Dear’s me one more time, I’ll scream. Yet we all sense that these quasi-words are special. Most speech is filled

with ordinary words like dog and walk that feel as if they conform to a set

of standards for a basic word in English. These words, when unadorned by

prefixes and suffixes, can be called canonical roots (roots for short), and

they are stored in a particular way in memory. A root occupies a distinct entry in the mental dictionary, like an entry in

a real dictionary. It specifies the word’s part-of-speech category, such as

“noun” or “verb.” It specifies the word’s meaning. And it specifies the

word’s sound, which conforms to a regulation template for standard words

in the language.11 In English the template for a canonical root is a

monosyllable, or a monosyllable with an unstressed bit dangling off the

end. Other languages have different templates; Italian, for example, does

not allow monosyllables for canonical nouns and verbs. (A rough and ready

test for a standard word sound in a language is the sound of its nicknames:

In English, uncanonical Bartholomew becomes canonical Bart; Elizabeth

becomes Lisa, Liza, Libby, Liddy, Lizzy, Liz, Betty, Betsy, Beth, or Bess. ) A

canonical root also embodies Ferdinand de Saussure’s conception of the

linguistic sign as an arbitrary pairing between a meaning and a sound, one

of the foundations of modern linguistics discussed in chapter 1.

12 Speakers tacitly sense that a canonical root doesn’t have to sound like its referent, as

does oink or pffffffffffffft; it symbolizes the referent by a conventional pairing they have learned. And here is the key to irregularity. An irregular plural or past-tense form

is a root linked to another root: sank to sink, feet to foot:

Irregulars by definition are arbitrary, and as we saw in chapter 3, they are

canonical English sounds: monosyllables such as stuck and mice,

monosyllabic roots adorned with prefixes such as became and understood,

or words with insubstantial second syllables such as children and oxen. The

fact that irregulars are tied to roots, not words, explains six cases in which

words cannot have irregular forms, even if their sound calls for one. The

explanation is that the words are not represented in the mind as canonical

roots, the only legitimate anchors for irregularity, but as stretches of sound

pressed into service as a word; the difference in mental representation from

the roots shown above might be as follows:

When a word is rootless and thereby disconnected from inflected forms

stored in memory, however, it is not left without a past tense or plural; the

rule rushes in and turns it into a regular form by adding a suffix. The first example of this effect is onomatopoeia, where a sequence of

vowels and consonants is construed not as a sound arbitrarily paired with a

meaning but as a direct rendering of a sound in the world. No one of course

thinks that onomatopoeic forms are particularly accurate, as in a tape

recording; they vary from language to language, have conventional forms,

and usually are compatible with the language’s sound pattern. But they can

violate canonical sound patterns—no word in English has a sequence like

oink—and crucially, people perceive them to resemble sounds. Onomatopoeic verbs and nouns need past tense and plural forms, but

because they are not canonical roots, they cannot tap into the lexicon of

roots and linked irregular forms that encourage irregular analogies. Onomatopoeic forms therefore are regular, even when their sound would

otherwise tempt people to borrow an irregular pattern, spling—splang—

splung style:

The engine pinged \[not pang or pung].

My grant got dinged \[not dang or dung].

That presentation really zinged \[not zang or zung].

The canary peeped \[not pept].

Her computer beeped \[not bept].

A second kind of sound that lacks a canonical root is a quotation. A

quotation does not have to use canonical words; it reflects a stretch of sound

that someone else has said, as in Elmer shouted, “Dwat!

”.

They may happen to be real words, of course, as in I hate how he begins every

sentence with “actually, ” but that is just a coincidence; you can quote any

sound anyone else has made. As with onomatopoeia, quotations are not

perceived as roots and fail to link to roots and their associations in memory. Stored irregular plurals are not tapped, and the regular applies, as in While

checking for sexist writing, I found three “man’s on page I (not three

“men” ).

In Jane Austen’s Mansfield Park, a character says of his drama-

loving father, “How many a time have we mourned over the dead body of

Julius Caesar, and to be’d and not to bed, in this very room, for his

amusement?”

A third way a word can be rootless is to be based on a name. In modern

English, names are meaningless noises. Many happen to sound like roots,

like Shepherd and Green, because surnames originally were based on a

person’s residence, occupation, father, or distinguishing features. But these

names have long since lost their meanings; no one expects Professor

Shepherd to be a shepherd or Mrs. Green to be green. Other names may

have nothing to do with English words, such as Dweezil Zappa, Carl

Yastrzemski, or Seamus McGillicuddy. As with onomatopoeia and

quotations, names are mentally registered as stretches of sound, not

canonical roots, and hence do not hook up with roots with the same or

similar sounds in memory—and that forces people to regularize them:

We’re having Julia Child and her husband over for dinner. You

know, the Childs are really great cooks \[not the Children].

Why hasn’t the German literary world seen any more Thomas

Manns \[not Menn]?

All the producers are looking for likable historians, but there aren’t

many Shelby Footes out there \[not Feete].

A fourth class of rootless words are foreign borrowings such as latke

and cappuccino, which are patently not canonical English words but sounds

taken from other languages. These words, despite their alienation from the

lexicon of English roots, happily receive regular inflection, as in lathes and

cappuccinos. They do so even when irregular patterns beckon. Despite the

widespread irregular pattern in thief—thieves, leaf—leaves, shelf—shelves,

and life—lives, nouns originally borrowed from French or German have regular plurals instead. From French we have beefs, chiefs, and gulfs, not

beeves, chieves, or gulves; from German we have fifes, not the fives and dra (fifes and drums) of Richard Lederers irregular-loving Farmer Pluribus

whom we met in chapter 3.13 If your pet mongoose gives birth, you have mongooses, not mongeese, because the word comes from mangus in

Marathi, a language of south India. More than one talisman is a bunch of

talismans, not talismen, because the source is the Arabic tilasm. Almost all

of the thousands of French and Latin verbs that were loaned to English

since 1066 are regular: derided, not derode; succumbed, not succame. Modern English speakers, of course, do not have a collective memory of

the cadences of an ancient Saxon fatherland. There must be a source in a

speaker’s own experience for the inkling that a word is not of native stock. Multilingual or cosmopolitan speakers may literally recognize a foreign

word, which is probably how chiefs, succumbed, and mongooses began

their English lives with regular plurals in centuries past. And if the first

speakers of a new word use a regular plural, other speakers generally follow

suit, because people pay attention to any irregular-sounding word that

appears in a regular form (as we saw on page 137).

But even monolinguals

can recognize loan words when they violate the canonical English sound

pattern. Recent immigrant words like cappuccino immediately give

themselves away, and even long-established French and Latin words have a

distinctive sound: They tend to be bisyllabic with stress on the second

syllable, such as deride. Even when speakers are unaware that a word at

some earlier point was a foreign borrowing, they may sense that the words

sound fancy or stuffy and not good old everyday English. (Several

experiments have found that people prefer native-sounding words to

Latinate-sounding words in a variety of everyday English constructions,

even if they cannot put their finger on the difference.14 ) Conversely, when

speakers have no sense at all that a word has been borrowed they treat it as

a standard root, and in those cases they feel free to make it irregular if it

resembles other irregular roots. Quit and cost, both imported from French

but assimilated as standard English monosyllables, are examples. Their

past-tense forms are irregular no-changers, analogous to hit and cast. A fifth class consists of words that are recognized as rootless because

they were concocted by artificial means. To synch is a truncation of to

synchronize, and its past tense is generally synched (as in lip-synched), not

sanch or sunch. The system manager of a computer installation is

sometimes called a sysman, plural sysmans. Acronyms are other examples

of ham-fisted wordsmithing, and they easily undergo regular suffixation, as

in PCs, TVs, SOBs, and the much-maligned RBIs. Even when an acronym

matches an irregular sound, the irregular form is unavailable and regular

suffixation applies. If a container with a mixture of oxygen and xenon were

labeled with the acronym OX, it is doubtful that several tanksful would be

called OXen. The example is a bit contrived, but in other languages they are

plentiful, as we shall see in chapter 8.

The sixth and final example consists of words that lack their own roots

because they are converted from a root of a different part-of-speech

category. English is notorious for converting roots from other categories

into verbs. Columnist John Powers wondered at the beginning of this

chapter who made impact a verb, and Calvin tells Hobbes:

CALVIN AND HOBBES © 1993 Watterson. Reprinted with permission of Universal Press Syndicate. All rights reserved. Far from weirding the language, verbed nouns are punctiliously lawful. Verbs that are recognized as thinly disguised nouns or adjectives don’t

accept irregular forms, even when they sound like an irregular verb:

Boom-Boom Geoffrion got high-

hit with a high stick

sticked!

\[not high-stuck]

Powell ringed the city with

formed a ring around

artillery. I steeled myself for a visit with

made like steel

my dentist, Dr. de Sade. Harvey hared his soul on Oprah. laid bare

Mongo spitted the pig. put on a spit

Vernon braked for the moose. applied the brakes

Mae Jemison out-Sally-Rided

outdid Sally Ride

Sally Ride. Swans are dark-meated fowl. having dark meat

Poor Bowser had to be de-flea’d. had fleas removed

Babs quickly righted the canoe. set right

We sleighed over the river and

went by sleigh

through the wood. Martina two-setted Chris. beat in two sets

After you’ve meaned both

computed the mean (average)15

columns, you can do the t-

test. Mom was flying home. In a box. given a wake 16

To be waked and buried. Most snow or sugar snap peas

have the string removed17

need to be stringed. The explanation is that a noun root like stick cannot have an irregular

past tense associated with it because the concept of past tense makes no

sense for a noun and hence cannot be listed with it. (What could it possibly

mean for hockey stick to have a “past tense”?) The irregular past-tense form

stuck that is stored with the verb root stick is not treated as relevant, because

to high-stick doesn’t have that root. The regular rule is not restricted to verb

roots or to anything else, but applies by default. The rule therefore is fully

available—indeed is the only way—to inflect verbs without verb roots. The

same thing happens with nouns (as we shall see in chapter 8 \[page 223]).

We almost have an explanation for flied out. Baseball fans recognize

that the verb to fly out is based on the noun a fly, a. k.a. a can of corn, namely a high, arcing ball. To fly out means to make an out by hitting a fly

that gets caught. One might conclude that flew out is avoided because it is

based on a noun, just as high-stuck and de-fled are avoided because they are

based on nouns. But we have a problem. The noun a fly was itself converted

from the everyday verb to fly, meaning “to slip the surly bonds of earth and

dance the skies on laughter-silvered wings.”18 Since the baseball fly is a double convert—a verb from a noun from a verb—it does have a verb root,

and that root comes with flew and flown. Something is missing from the

explanation. The missing piece comes not from the nature of words but

from the nature of rules. The point of grammatical rules is to define new combinations in which the

meaning of the whole can be computed from the meanings of the parts and

the way they are arranged. Some rules (the rules of syntax) build sentences

and phrases out of words; others (the rules of morphology) build complex

words out of simple words and bits of words such as prefixes and suffixes. When a new word surfaces, like weirding or long-hilled thrasher,

grammatical rules allow the speaker to coin it and the listener to understand

it. Take the verb overeat. It is based on the verb root eat:

The root is then encrusted with a prefix, yielding the following mental

structure, in which the V at the top stands for the whole word overeat:

How do we know how to use the new word overeat?

Easy—we give it the

properties of the rightmost thing inside it, eat. What part of speech is

overeat?

It is a verb, just as eat is a verb. What does overeat mean?

It refers to a kind of eating—eating too much—just as eat refers to eating. And what

is its past-tense form?

Overate, not overeated, just as the past-tense form of

eat is ate, not eated. A new complex word inherits its traits (including any irregular forms)

from the memory entry for the rightmost word inside it, the head of the

word. The pipeline of information from the head at the bottom of the tree to

a new, complex word formed from it, and then to an even bigger word

formed from that one, can be depicted like this:

The percolation of information up from the head can explain the

examples that disconfirm the semantic stretch theory (the theory that a

change of meaning poisons an irregular).

For example, the compound

workman is formed by prefixing the noun man with the verb work:

The traits of the whole word come from the traits of the rightmost word

inside it, the head, in this case man. Workman is a noun, just as man is a

noun. It refers to a kind of man, just as man refers to a man. And its plural

is workmen, because the plural of man is men. A similar explanation works

for other compounds and metaphors, such as bogeymen, superwomen, and

sawteeth, and for prefixed verbs such as understand—understood and

become—became. It also works for idioms like took a nap and threw up, as

soon as we remember that these are phrases, not words, and that in English

the head of a phrase is on the left, not the right (hence the great mothers-in-

law/mother-in-laws debate in chapter 2).

All of which brings us to lowlifes, flying out, and Mickey Mouse. A

few complex words are headless: They can’t get their properties from their

rightmost element if they are to work the way the speaker wants them to

work. The information percolator must be turned off for the word to be

interpreted and used properly. As a result, the pipeline that carries stored

information from the word’s root is clogged, and any irregular form stored

with the root is imprisoned in memory, unable to percolate up to apply to

the whole word. The regular rule, acting as the default, steps in to supply

the word with a past-tense form, undeterred by the fact that the sound of the

word smacks of irregularity. How does a word lose its head?

One way is to be a compound that

doesn’t refer to the kind of thing indicated by its rightmost word. Instead it

refers to something else, which merely has or does something to the kind of

thing indicated by the rightmost word. Though a workman is a kind of man

and a bluebird is a kind of bird, a cutthroat is not a kind of throat, nor is a

lazybones a kind of bones. Linguists call these bahuvrihi compounds, from

the Sanskrit expression “having much rice.”19 (The term comes from a

school of Indian linguists working 2500 years ago who left us a remarkably

sophisticated analysis of the grammar of Sanskrit.) Similarly, a lowlife is

not a kind of life, but a kind of person: a person who has (or leads) a low

life. For it to have that meaning, the percolation pipeline, which would

ordinarily make lowlife mean a kind of life, must be plugged up. With the

data pipeline to memory disabled, there is no way for the other information

stored with life to be passed upward either, such as the linked plural form

lives. With the irregular plural unavailable, regular -s gets the call, and we

have lowlifes. Headlessness explains at least four kinds of plurals and past-tense forms

that have puzzled word-watchers for decades. Bernstein complains that “a

few plurals seem almost unreasonable: talismans, mongooses, still lifes.” 20

We have already demystified the first two, and now we see that a still life is

not a kind of life but a kind of painting. To have that meaning it must be

percolation-proof, sealing off lives in the mental lexicon and defaulting to

lifes. Similarly, a flatfoot is not a kind of foot but a neighborhood policeman

(his feet flattened from so much walking), and several of them are called

flatfoots, not flatfeet. An inexperienced woodsman or cub scout is called a

tenderfoot, and dictionaries give tenderfoots as one of its plurals. If I were Professor Kugelmass in the Woody Allen story about the

machine that could project people into the novel of their choice, I would ask

to be projected into Richard Russo’s Nobody’s Fool so I could resolve a

controversy at the local high school:...

A controversy had erupted on the editorial page of the North

Bath Weekly Journal over whether the plural of Sabertooth should

be Sabertooths or Saber-teeth. When the cheerleaders led the spell

cheer, how should it go?

The principal said Saberteeth sounded

elitist and silly and dental. The chair of the high school’s English

department disagreed, claiming this latest outrage was yet another

symptom of the erosion of the English language, and he threatened

to resign if he and his staff were expected to sanction tooths as the

plural of tooth. Why not?

the public librarian had asked in the next

letter to the editor. Wasn’t this, after all, the same English

department that had sanctioned “antelopes” as the plural of

“antelope”?

The letters continued to pour in for weeks. Beryl

Peoples, who’d nursed a twenty-year grudge against the principal

for caving in and allowing history courses in the junior and senior

high school to be redesignated “social studies,” had the last editorial

word, reminding her fellow citizens that the sabertooth tiger was an

extinct animal. Food, she suggested, lor thought.21

Fortunately, they worked it out without my help: The new banner read “GO

SABERTOOTHS!

TROUNCE SCHUYLER SPRINGS!” It was the right

decision, because sabertooth is a bahuvrihi: It refers not to a kind of tooth

but to a kind of cat. I would also defend Bilbo Baggins in J. R. R. Tolkien’s The Fellowship

of the Ring:

“My dear Bagginses and Boffins,” he began again; “and my dear

Tooks and Brandybucks, and Grubbs, and Chubbs, and Burrowses,

and Hornblowers, and Bolgers, Gracegirdles, Goodbodies,

Brockhouses and Proudfoots.” “Proud-FEET!” shouted an elderly

hobbit from the back of the pavilion. His name, of course, was

Proudfoot, and well merited; his feet were large, exceptionally furry,

and both were on the table. “Proudfoots,” repeated Bilbo.22

And I wish I could have put my two cents into this conversation between a

know-it-all narrator and a character called Buffalo Gal in a short story by

Alison Baker:

“They come and they go,” Buffalo Gal said. “They might as well be

bigfoots.”

“Bigfeet,” I said. “Whatever,” Buffalo Gal said. 23

Here is another mystery solved by the word structure theory. A 1989

article in Newsweek began:

It’s been ten years since the Sony Walkman was born. Fifty million

of the machines have been sold. Yet nobody knows the correct

plural form of Walkman. Is it Walkmans?

Is it Walkmen?

We can

only guess. Nonsexists might suggest a new name: Walkperson. But

then, would the plural be Walkpersons or Walkpeople?

Sony Corp. avoids the issue entirely by using Walkman only as an adjective. In

the interest of consistent usage—and trademark protection—Sony

talks about “Walkman® personal stereos.” Everyone else, when

talking about personal stereos—whether Panasonic, Toshiba or

Aiwa—calls them “Walkmans.” Or “Walkmen.” Never “personal

stereos.”24

Some people, at least, are completely confident that the plural should be

regular. The owner of a San Francisco store had it bent into the tubing of a

neon sign:

A walkman, of course, is not a kind of man, so many of us, like the sign

maker, interpret it as headless and hence bereft of the irregular plural men. It’s not exactly a bahuvrihi either, because while a lowlife has a low life and

a flat-foot has flat feet, in no sense does a walkman have a man. The closest

gloss might be “that which allows a man to walk (while listening to

music).” Parsing it may be futile, however, because Japanese companies

often use meaningless English names and slogans just for the cachet, such

as “Supreme Liberal,” “For vibratory refreshment,” and “Love the earth

with honest poverty.”

Headlessness also explains a second class of regularizations, eponyms. An eponym is a word that comes from a name, such as atlas, boycott,

bowdlerize, cinderella, maverick, quixotic, sandwich, scrooge, shylock,

tantalize, and, according to legend, c rap, after Thomas Crapper, a

nineteenth-century British inventor who improved the flush toilet. (Crapper

really existed, but the noun, originally “chaff,” dates back to Middle

English.)25

A Mickey Mouse, in the sense of a simpleton, is an eponym. It began

with the ordinary noun mouse. Walt Disney made it a name when he

christened his diminutive hero Mickey Mouse. Names are somewhat like

nouns, but in English they are not the same thing (you don’t ordinarily say

She’s talking to the Mildred or I left work because sick Jason came home

early).

Then in colloquial speech the name was converted back into a

common noun, a Mickey Mouse:

The new noun is headless, because the percolation pipeline had to be

blocked twice: once to convert the noun mouse into a name, and then to

convert the name back into a noun. (It also had to be blocked to get the

meaning to come out right: A Mickey Mouse is not a kind of mouse in the

sense that a workman is a kind of man.) With percolation turned off, mice is

trapped in the lexicon, and the plural rule gives us Mickey Mouses. Note

that Mickey Mouses, a double convert, gets a slightly different explanation

from the Childs, a single convert. With Mickey, but not with Julia, the

surname has a visible connection with the ordinary noun, so the problem is

not the lack of a root; it is the inaccessibility of a root. Nouns can be based on other kinds of names, such as works of art,

products, or teams. We might say that Michael Keaton starred in the first

two Batmans, not the first two Batmen. I can imagine someone arguing that

Roy Orbison’s original recording is the best of all the Pretty Womans, and

Bobby Darin’s is the best of all the Mack the Knifes. I have seen or heard

mentions of Spectrums and Quantums (bicycles), Elfs (cars), John Deeres (tractors), Top Shelfs (frozen dinners), Sea Wolfs (navy Aircraft), Supermans

(comic books), and Maple Leafs (gold coins).26 In Popular Photography a journalist wrote of a new camera, “As Canon squeezes out more production,

ELPHs keep selling out. (What is the plural of ELPH, anyway?

ELPHS?

ELVES?

ELVIS?)”27 As for sports teams, we have the Maple Leafs high-

sticking in Toronto and the Marlins (not Marlin) hoisting cans of corn in

Florida. At this point killjoys will bring up the Timberwolves, who shoot

hoops in Minnesota; I will get to them, and to other apparent

counterexamples, in the next section. At long last, the complete explanation for why no mere mortal has ever

flown out to center field. Recall that in the evolution of baseball argot the

plain verb to fly was converted to a noun, a fly, which was then converted

back to the verb to fly, meaning “to hit a fly that is caught”:

The new verb (top V) is sealed off from the root verb (bottom V) at two

layers, the one that converted the verb to a noun, and the one that converted

the noun back into a verb. Percolation had to be blocked both times to allow

the verb to change categories rather than blindly receiving the category

from one layer down. Baseball cognoscenti can hear the fly ball in flying

out, so for them the forms flew and flown are unable to climb out of the lexical entry for fly. The word turns to -ed as the last resort and becomes

flied out. Here are some other verbs-from-nouns-from-verbs that shook off the

irregularity of their roots:

Once again, Perot grandstanded to the audience. (to stand → a grandstand → to play to the grandstand)

Vera costed out the grant. (to cost → the cost → to ascertain the cost)

Doctor Crunch encasted my leg. (to cast → a cast → to put in a cast)

She threw out all her runned nylons. (to run → a run → having a run)

A doctor who slided a sample. 28

(to slide → a slide → to place on a slide)

Many projects could be offshooted from television in the

classroom. 29

(to shoot → an offshoot → to make an offshoot)

In each of the past two seasons, Cleveland State guard William

Stanley has sported a self-styled, one-of-a-kind hairdo. In

1987–88 it was a half-foot-high flattop. Last season he went to

a bilevel box cut. This season, as a senior, Stanley has outdo’ed

himself. 30

(to do hair → a hairdo → a ’do → to have a more impressive do

than)

The preterite of to joyride is not joyrode, nor even joyridden,

but joyrided. 31

(to ride → a joyride → to take a joyride)

Headlessness explains a fourth curiosity, this one a quirk of spelling. English spelling is a rule system that connects the sounds of words with

their written forms. As with grammar, spelling is rife with irregularities and

complications, especially in frequent words such as eye, of, have, and

would. One complication is that when a noun ending in y gets a suffix, the y

becomes ie, as in the plurals army—armies, body—bodies, and cherry—

cherries, and the derived nouns in happy—happiness, pretty—prettiness,

and ugly—ugliness. Yet the regular spelling rule, in which an ē sound at the

end of a stem is spelled y, sometimes reasserts itself. Here is an example

from an article on fashions in women’s names:

Bettys abound, past and present—from Crocker to Boop, from

Grable to Friedan to Ford....

If you’re looking for a Betty under

40, though, good luck....

Bettys are so endangered that they’ve

formed a club—lots of clubs, actually. In fact, the Bettys of

Nebraska just held a convention simply to rejoice in their

Bettyness.32

Betty, of course, is not just any old noun but a noun that comes from a

name. And pluralized names shed their irregular spellings, just as pluralized

names shed their irregular plurals: the Kennedys, the Fogartys, the Kansas

Citys, the Germanys, the Emmys, the Tonys; I’ll have two Bloody Marys. So

do nouns converted from other grammatical categories: A sign outside an

apartment-motel advertised Dailys Weeklys Monthlys Yearlys, all nouns

derived from adverbs. 33 When a noun is based on a nonnoun, people seal

off their associations to irregularities, even in spelling. The spelling effect doesn’t always work. I have seen Dollies (the

eponymous cloned sheep), dailies (newspapers), onlies (only children), and

goody-goodies, not to mention the Alleghenies and the Rockies. Nevertheless it is striking that spelling, which most of the time is a clumsy

afterthought pasted on to the standard equipment of language, should often

showcase a deep principle of grammatical organization. An irony of systematic regularization is that ordinary speakers apply

abstract grammatical principles instinctively, while many style manual

authors and language mavens are oblivious to them and hector people into

sticking with the irregular. Here is Theodore Bernstein contrasting flied

(which he explains with the semantic stretch theory) with broadcasted:

If you think you have correctly forecasted the immediate future of

English and have casted your lot with the permissivists, you may be

receptive to broadcasted, at least in radio usage, as are some

dictionaries. The rest of us, however, will decide that no matter how

desirable it may be to convert all irregular verbs into regular ones,

this cannot be done by ukase, nor can it be accomplished overnight. We shall continue to use broadcast as the past tense and participle,

feeling that there is no reason for broadcasted other than one of

analogy or consistency or logic, which the permissivists themselves

so often scorn. Nor is this position inconsistent with our position on

flied, the baseball term, which has a real reason for being. The fact

—the inescapable fact—is that there are some irregular verbs.34

This had long been a losing battle. In The American Language Mencken

notes that “The effort of purists to establish broadcast as the preterite has

had some success on higher levels, but very little on lower. ‘Ed Wynn

broadcasted last night’ is what one commonly hears.”35 Henry Fowler, like

Bernstein, was scornful, though his rationale is closer to the linguistic truth:

If etymology is to be our guide, the question whether we are to say

forecast or forecasted in the past tense and participle depends on

whether we regard the verb or the noun as the original from which

the other is formed. If the verb is original (= to guess beforehand)

the past and p. p. \[perfect participle] will be cast as it is in that verb

uncompounded; if the verb is derived (= to make a forecast) they

will be forecasted, the ordinary inflexion of a verb. The verb is in

fact recorded 150 years earlier than the noun, and we may therefore

thankfully rid ourselves of the ugly forecasted; it may be hoped that

we should do so even if history were against us, but this time it is

kind. The same is true of broadcast; and broadcasted, though

dubiously recognized in the OED Supp., may be allowed to die.36

If only Fowler had couched his explanation in terms of speakers’ intuitions

about whether the verb was based on the noun or vice-versa, rather than on

whether the verb or noun appeared in the language first (a bit of history lost

on modern speakers), he would have been less offended. People may very

well interpret the verb to broadcast as “to do a broadcast,” rather than

interpreting a broadcast as “an act of broadcasting”; in contemporary

English the noun is far more frequent than the verb and may feel more

basic. 37

Other indignant critics also forget to ask whether a noun may be lurking

behind a supposedly slovenly use of a regularized verb. The Boston Globe

ombudsman spinelessly agreed with a reader’s complaint about one of his

colleagues:

A woman wrote: “I join other readers in lamenting the lack of

attention given to good writing, spelling, and grammar these days.”

One article she sent left out a key comma and contained the phrase

“he may of been.” Another article read, “Martyny subletted a

Kenmore square apartment.” It’s sublet.38

Not necessarily!

In American English hardly anyone uses the verb to let,

meaning “to lease,” but many speakers use the noun a sublet. They might

analyze to sublet as a verb from a noun (to arrange a sublet) rather than as

the verb to let with the prefix sub-.

And that would lead to a regular past-

tense form. Similarly ashamed, Scientific American published this letter

under the heading “Nobody’s Perfekt”:

Have you actually used “inputted” as the past tense of a verb?

Yes,

in the caption of the figure on page 150 of the January issue. I am

upset. I am appalled. I am horrified. I am out putted.39

The nouns input and output are about sixty times more common than the

verbs to input and to output. If the caption writer interpreted to output as “to produce output,” the regular past-tense form would follow as night follows

day. Before we congratulate the word structure theory for explaining so many

kinks and vagaries, we need to worry about the examples it doesn’t explain. I’ve already mentioned the Timberwolves and the factions that like Bigfeet

and Saber-teeth, and I confess to having heard flew out and flown out a number of times. A skeptic might wonder whether regularization is really

that systematic after all. Perhaps people carelessly turn irregulars into

regulars, just as they have been doing with simple verbs for hundreds of

years ( chid becoming chided, holp becoming helped, and so on).

Perhaps I

have cherry-picked a few examples that are consistent with a needlessly

fancy theory. How can we prove that these effects are alive in the minds of speakers?

By bringing the speakers into the lab, presenting them with new headless or

rootless verbs, and seeing whether they regularize them more than they

regularize pure verb roots. John Kim, Alan Prince, Sandeep Prasada, and I

gave people a questionnaire with three dozen irregular-sounding verbs. 40

Half the sentences used the verb root metaphorically:

When guests come, if they arrive with slides my hopes for a lively

evening quickly sink. When I saw Bob and Margaret carrying six boxes, my hopes

sinked instantly. When I saw Bob and Margaret carrying six boxes, my hopes

sank instantly. These sentences served as a control group. They contained a verb in a

metaphorical or extended sense, so the semantic stretch theory predicted

that they would scare people away from the irregular forms. We were

confident, though, that people would stick with the irregular forms for these

items, because they have the same root as the standard irregular verbs. So

we used the ratings of these items as a baseline measure of people’s

fondness for the various irregular past-tense forms. The sentences we really were interested in contained verbs based on

nouns:

When guests come, I hide the dirty dishes by putting them in boxes

or in the empty sink. Bob and Margaret were early so I quickly boxed the plates and sinked

the glasses. Bob and Margaret were early so I quickly boxed the plates and sank

the glasses. If the word structure theory is true and people spontaneously regularize

headless and rootless words, these verbs-from-nouns should be rated

differently from the controls: People should flip to regular sinked, or at least

should dilute their preference for irregular sank. We matched up the items

in pairs, every stretched verb having a verb-from-noun with the same sound

and vice-versa. Across the experiment the matched sentences were

presented equally often so that any fondness or distaste for an irregular

sound would cancel out. Every subject, though, saw a particular verb such

as sink only as a stretched verb or only as a verb-from-noun, so they

wouldn’t make side-by-side comparisons and concoct their own theories of

what we were after. We also tested a few verbs-from-nouns that are already

in the language: flied out, grandstanded, ringed the city, and so on. The results were gratifying. With every verb, the subjects liked the

regular past tense (such as sinked) better when presented as a verb-from-

noun ( sinked the dishes) than when presented as a stretched verb ( my hopes

sinked).

For 90 percent of the verbs they liked the regular form so much that

they rated it higher than the irregular form—including, to our relief, the

item flied out. We also tested the semantic stretch theory, asking a second

group of subjects to rate the degree of stretching in the meanings of the

verbs (for example, how much the verb to sink has been stretched in our

hopes sank).

If the theory is correct, those ratings should predict the

preferences of the original group of subjects for regular and irregular forms. Using a standard statistical procedure, we pitted semantic stretching against

word structure to see whether one, the other, or both were necessary to

explain the preferences for regularity or irregularity. Word structure was

necessary to explain the preferences; semantic stretching was unnecessary. We also wondered whether the effect might be a fussy affectation of

pointy-headed, Volvo-driving, endive-nibbling, chablis-sipping young urban

professionals. It seemed unlikely, given that most language mavens fail to

grasp the principle, but still worth ruling out. So we gave the questionnaire

to a new sample of people without a college education, recruited through an

ad in the local tabloid paper. The results were the same. Why then would Chris Berman of ESPN say that “Jose Offerman flew out

to center field in the ninth inning”?

41 Why would William Safire, of all people, refer to the president’s spin doctors as “the bigfeet of the Opinion

Mafia”?42 Why is there a Native American nation called The Blackfeet, and

a species of goose called pinkfeet, when they should behave like Proudfoots

and flatfoots?

And what about those damn Timberwolves?

Now that we have the results of the experiment we needn’t worry about

the phenomenon itself; people do in fact regularize rootless and headless

words. At worst, the exceptions force us to say that regularization is a real

effect but only a statistical one. It tips the odds away from the massive

tendency to avoid flied and foots and wolfs in ordinary speech, even if it doesn’t flip the preference all the way to 100 percent use of the regular form

all of the time. But the word structure theory is in even better shape than that. Like its

parent, the words-and-rules theory, it is about the psychology of flesh-and-

blood speakers rather than some scholar informed by the best etymologies

philology has to offer. People should regularize headless forms only when

they perceive the words to be headless. They may not be conscious of a

word’s derivation, or be able to explain it to others, but they should have a

sense that the word is based on another word (for example, that to fly out is

based on a fly).

When they don’t—when they are oblivious to the noun in a

verb-from-a-noun and imagine that it is just a stretched verb root—the

theory predicts that they should stick with the irregular. First let’s consider

some ways in which people might misanalyze headless words; then I’ll

present evidence that whenever people do so, they stick with irregular

forms. Sometimes there is a way to stretch a verb root to refer directly to an

action, skipping the middleman noun. In sportscasting it’s common to

personify the ball and describe it with the name of the player who propelled

it and whose fate is tied to it. Basketball commentators often say Jordan got

blocked and Larry is rejected, where it is the ball that is blocked or rejected,

not the man. If an announcer similarly personifies a flying baseball as its

hitter, the hitter could be said to have flown out; the noun fly need never have entered his mind. A bahuvrihi noun also can invite misanalysis, because people may think

not of the possessor of the rightmost word but of the rightmost word itself. In the rhetorical device called synechdoche, a part symbolizes the whole, as

in The Celtics need fresh legs; She got a new set of wheels; I counted heads;

and He is chasing skirts. Perhaps Blackfeet, pinkfeet, and the occasional

bigfeet and ten-derfeet are being used synechdochically as standard

compounds like fresh legs rather than literally as bahuvrihi compounds like

lowlife. In naming a sports team (or understanding its name), usually its

members are metaphorically identified with some referent (a Lion, a Tiger,

a Bear), and the name of the referent is then pluralized (The Lions, The

Tigers, The Bears).

Pluralized names, as we saw with the Childs, the

Manns, and the Mickey Mouses, lose their irregularity. So when a Maple

Leaf joins his teammates on the ice, they are the Maple Leafs. But there are other ways to name a team. Instead of metaphorically

naming a member and pluralizing the name, one can name the entire team

at once. That’s what gave us those tacky mass-noun teams: The Utah Jazz,

Miami Heat, Orlando Magic, Colorado Avalanche, Tampa Bay Lightning,

Dallas Burn, San Jose Clash, and Kansas City Wiz. Entire teams also have

been given x-rated plural names such as the Boston Red Sox, Chicago

White Sox, Everett Aquasox, and West Tenn Diamond Jaxx. No singular

was ever pluralized; no player identifies himself as a “Red Sock.”

Now, if entire teams can be identified with mass or plural referents, it’s

easy to imagine that the roundball team in the Land of Lakes is being

identified with a pack of wolves—wolves do, after all, fast-break in packs. No one ever had to figure out how to pluralize a Timberwolf because the

whole team was named after the plural timberwolves to start with. (Once the

team is named, it’s easy to work backward to the name of a single player, a

Timbenvolf. ) The situation is different north of the border, where

Torontonians began with the symbol of Canada, the proudly singular Maple

Leaf. So there are many ways in which people may at times shortcut a

derivation. But without independent evidence that people have taken the

shortcuts exactly when they fail to regularize, the explanations would

merely be escape hatches for the word structure theory, making it

unfalsifiable. So Kim, Prince, Prasada, and I went back to the lab to break

the logical circle. We wanted to measure people’s tendency to shortcut a

derivation, or to encourage them to do so, and then see if that increased the

appeal of irregular forms.43

First we dusted off our materials from the earlier experiment and asked

a new group of subjects to rate, on a scale of 1 to 7, how similar in meaning

the noun was to the similar-sounding root verb. We asked our subjects,

“How similar is the fly in fly ball to th e fly in birds fly south?

How similar is

the sink in kitchen sink to the sink in sink the ship?” and so on. Their responses gave us a measure of how vulnerable each verb-from-noun was to

being misanalyzed: With nouns that feel similar to the soundalike verbs,

people might be more tempted to misanalyze the verb-from-noun as a

simple variant of the verb. (For example, if a fly ball was perceived as

being similar to what birds do, people might misanalyze fly out as a

symbolic version of to fly rather than as a conversion from a fly. ) That in

turn should lead to seemingly embarrassing irregular forms such as flew and

flown. Note the contrast with the semantic stretch theory: We are predicting

that a metaphoric interpretation should make the irregular more appealing,

not less so. With these ratings in hand we went back to the data from the original

experiment. As predicted, the verbs-from-nouns that were rated as most

shortcut-table were the ones for which the earlier subjects had been least

attracted to a regular form such as flied out. In another experiment we gave people irregular-sounding verbs and

nouns that were clearly related to one another, and tried to manipulate

whether the verb was perceived to be based on the noun or the noun was

perceived to be based on the verb. Compare these two kleeds:

Mary got a brand new kleed for her birthday. She liked it so much, she kleeded/kled for a week. It has been a long time since I have had a nice, long kleed. I kleeded/kled quite often in the old days. In both cases people see a noun and then a related verb. In one, the noun

refers to a physical object and the verb is based on it: to kleed means to use

a kleed. In the other, the verb is an action and the noun is based on it; a

kleed is an interlude of kleeding. The stimuli were identical; only the

readers’ analysis of the verb—as a verb-from-a-noun or as a root verb—

varied. As we predicted, that difference affected their ratings of the past-

tense forms. With a verb-from-a-noun (use a kleed), regular kleeded was as

acceptable as irregular kled; with a simple verb that just happened to be

accompanied by a noun-from-a-verb (have a long kleed), the regular form

was much less acceptable than the irregular. 44

Both experiments show that the exceptions to the regularization effect

are exceptions that prove the rule: When people don’t perceive a word as

headless, they don’t plug the pipeline that sends irregular forms up from

memory either. That secures a major kind of evidence for the word structure

theory and for its parent, the words-and-rules theory. Irregular forms are

word roots stored in memory; regular forms are computed when memory

fails to cough up a form, for any reason. That in turn shows that a rule is a

mental operation that manipulates variables, such as “verb” or “noun,”

rather than an association to concrete memories of particular words and

their sound patterns. It also shows that people erect an abstract mental

scaffolding around words. The memory blockage in the examples in this

chapter come from the nature of the mental scaffolding: People sense

whether a word is stored as a root, and whether a word has a structure that

allows information about the root to percolate up from memory. When I lecture on regularity and irregularity in language, the question I am

asked most often is, “What’s the deal with the plural of computer mouse?”

Here, as a public service, is my best guess. The use of mouse for “pointing device” goes back to 1965.45 The first

mouse had its wire coming out the front toward the user, and it reminded

the inventor, the computer scientist Douglas Engelbardt, of Mus musculus. Decades later the wireless pointing device was introduced; it is sometimes

called a hamster. Many people are squeamish about referring to more than one of them as

mice. In 1992 I surveyed several dozen mail-order ads and found that many

used plural headings for every category of hardware but the mouse, like

this: Desktops—Notebooks—Monitors—Printers—Keyboards—Mouse. A

few others played it safe by advertising Pointing Devices or Input Devices. More than half did use mice; none used mouses. Micephobia is beginning to

abate, and today mice is common in stores, magazines, and web pages (for

example,

my

local

CompUSA

has

an

aisle

labeled

“KEYBOARDS/MICE”).

Many people, though, still wince at mice, though

not to the point of using mouses, which remains rare among native

speakers. I would love to chalk up mice avoidance as another case cracked by the

word structure theory, but in this case it is of no help. The root of mouse the

pointing device is indubitably mouse the rodent, and the word is based on a

transparent metaphor that should allow the irregular plural to bubble up

unscathed. Thankfully, the facts don’t call for such heavy-duty machinery in

any case. People are a bit skittish about mice; they don’t shun it entirely or

switch to mouses. We need a different kind of explanation. The explanation comes in two parts. One is familiar: Irregular forms are

stored in memory, regular forms don’t have to be. The other concerns the

nature of the concept “plural,” which is not as straightforward as one might

think. Plural means “more than one of,” but there are lots of ways in which

objects can come in multitudes. We can behold a small number of objects, each apprehended as an

individual:

Several of them can be parts of a larger object:

Each can be a part of a larger object, several of which are under

consideration:

The objects can congregate in an amorphous swarm or mass:

Or they can be distributed diffusely and randomly in the surrounding

environment:

Many languages don’t even have a unified plural marker that treats all the

more-than-ones the same; they use various constructions for pairs, swarms,

herds, families, and so on.46

Suppose the regular plural suffix -s simply means “more than one of,”

so that hands = “more than one” + “hand” and rats = “more than one” +

“rat.” But suppose that no single concept of plurality is shared by all the

irregular plurals. They have to be stored separately in memory anyway

because of their idiosyncratic sounds, and that means each can have its own

meaning slot in which a unique, concrete representation of more-than-one-

of-that-kind-of-thing can be entered. It could even be a mental image of a

typical multitude of that kind of thing: a committee of men or women, a

flock of geese, a pair of feet, a set of teeth, a brood of children, a team of

oxen, an infestation of lice. Consider now what happens when you are called on to refer to more

than one pointing device. Pointing devices come one to a computer, and

several of them would imply several computers. But several little rodents

tend to scurry, unattached, throughout the house or in meadows and woods. The metaphorical aptness of “mouse: the single rodent” for “mouse: the

single pointing device” evaporates when we now have to think of “mice:

the scattered vermin” as a metaphor for “mice: the accessory attached to

each of several computers.” And that, I submit, makes people uneasy about

calling the pointing devices mice. Evidence?

First, the same thing happens with other nouns. Remember

that irregular plurals happily lend themselves to metaphors such as

sawteeth, God’s children, chessmen, and oil mice. But that happens only

when the kind of plurality of the original word (small set, swarm, pair,

attachment, and so on) matches the kind of plurality of the metaphor. When

it doesn’t, we get the same feeling of queasiness that surrounds computer

mice. For example, foot is often used as a metaphor for a nether extremity:

There was a cottage at the foot of every mountain. An ambassador was seated at the foot of each table. The page number is printed at the foot of each page. But feet come two to a body, and when a metaphorical foot comes one to an

object, the plurals are tainted:

There were cottages at the feet of the mountains. Ambassadors were seated at the feet of the tables. Page numbers are printed at the feet of the pages. Note that as with the pointing device, the odor surrounding the irregular

is not bad enough to drive people to the regular foots, though it sometimes

forces them to an awkward singular. According to the Encyclopedia

Britannica, the painter Paul Klee described his boyhood education as

“mountains immeasurably high but with no foot.” (“Mountains with no

feet,” though more accurate, does not sound right.) A similar mismatch

explains the slight strangeness of Parish and McHale had excellent first

halves, where each player had an excellent first half. A headline in the New

York Times about tastes in classical music read, “Classical Radio Plays Only

to Sweet Tooths.” Presumably it is because every listener has a single sweet

tooth, and sweet teeth would connote a mouthful of them. The plural-

mismatch effect may even have contributed to the Toronto Maple Leafs, a

collection of individuals quite unlike a mass of foliage. A disfavored person is sometimes compared to an ignominious animal:

Silly goose!

Clumsy ox!

Filthy louse!

When the animals congregate in flocks, teams, or infestations, though, it’s

strange to refer to several such people with the irregular plurals:

Silly geese!

Clumsy oxen!

Filthy lice!

Hence the lyrics from Gentlemen Prefer Blondes:

He’s your guy when stocks are high

But beware when they start to descend. ’Cause that’s when those louses

Go back to their spouses. Diamonds are a girl’s best friend.47

The choice of plurals therefore depends on how the mind construes

multitudes. That part of our cognitive machinery not only affects how we

use the language today; it can shape a language over centuries. Often it’s unclear whether a multitude is best perceived as one big thing

or many little things, as we see in expressions like “He can’t see the forest

for the trees” and “The whole is more than the sum of the parts.” Whenever

a collection of individual things is reconceptualized as a single gestalt, the

plural for that collection can cease to feel like a plural, and the language can

change. In his song One Hippopotami, the comedian Alan Sherman sang, “The

plural of ‘half’ is ‘whole’; the plural of ‘two minks’ is ‘one mink stole.’” It

is an astute observation. The linguist Peter Tiersma has found that whenever

a set of objects can easily be construed as a single assemblage, a regular

plural is in danger of congealing into a mass noun or an irregular plural. 48

This is happening today to the noun data, which often refers to large

quantities of information and which is easily conceived of as stuff rather

than things; the word is turning from a plural ( many data) to a mass noun

( much data).

The effect is widespread. In language after language things

that come in groups, such as children, gregarious animals, and paired or

clustered body parts, end up unmarked, irregular, or transformed into a

singular, sometimes to get pluralized all over again by a subsequent

generation of speakers. Nonstandard dialects are filled with double plurals

such as oxens, dices, lices, and feets, and that is how we got the strangest

plural in Standard English, children. Once it was childer, with the old plural

suffix -er also seen in the German equivalent Kinder. But people stopped

hearing it as a plural, and when they had to refer to more than one child,

they added a second plural marker, -en. Today many rural and foreign

speakers still don’t think of children as plural, and have added a third suffix,

yielding the triply plural childrens. A regular rule is a powerful instrument, creating inflected forms for a

motley collection of rare, strange, and eccentric words. Is there any place it

cannot work?

Indeed there is, and it is my final demonstration of the

difference in kind between regular and irregular inflection. Regular plurals don’t like appearing inside compounds. We speak of

anteaters, bird-watchers, Beatle records, Yankee fans, two-pound bags,

three-week vacations, and all-season tires, even though it’s ants that are

eaten, birds that are watched, all four Beatles that played on Sgt. Pepper’s

and the white album, and so on. 49 The discomfort is not shared by irregular

plurals, though, as we see when we lay compounds with regulars and

irregulars side by side. An apartment infested with mice is mice-infested,

but an apartment infested with rats is not rats-infested; it is rat-infested, even though by definition a single rat is not an infestation. Mice and rats are

similar creatures, so the effect is not caused by a difference in meaning; it is

caused by sheer irregularity. We also have teethmarks but not clawsmarks,

men-bashing but not guys-bashing, and a song about a purple-people-eater,

but never a song about a purple-babies-eater.50

Here are some real-life examples in which mice rush in where rats (or

trackballs) fear to tread:

Mice Bait (a sign outside a general store)

Mice Cube (a better mousetrap)

mice-drivers (Microsoft software)

I felt mice-feet of apprehension scurrying over my skin. (From The

Edible Woman, by Margaret Atwood)

Bad maps, mice-infested lodgings, and strict rules. (Description of

the Appalachian Trail in the New York Times Book Review)

Frozen Mice Sperm (headline)51

Cells Implanted in Mice Brains; Hope Is Voiced for Mental Ills

(headline)52

Mobile Phone Radiation Mice Tumor Link Much Stronger Than

Expected (headline)53

Mice Accessories (a sign in a computer store)

Not far from where I work a flock of geese has taken up residence, and the

city thoughtfully put up a sign declaring that part of Memorial Drive a

GEESE CROSSING. Had it been a flock of ducks, I doubt the sign would

have announced a DUCKS CROSSING. A store in Florida selling exotic

leather fashion accessories had a sign for “Chicken Feet” wallets (irregular

and plural) and a sign next to it for “Turkey Leg” wallets (regular and

singular).

A periodical called Rural Heritage describes itself as “a

bimonthly journal for small farmers and loggers who use draft horse, mule,

and oxen power.”

More examples can be found in the two-headed compounds that

Sanskrit grammarians called dvandva (two and two), in which two nouns

apply equally to some chimerical or twice-described person. 54 Examples

include man-child, manfish, man Friday, manservant, man-woman, woman-

doctor, girlfriend, boyfriend, boy-king, player-coach, and singer-songwriter. Dvandva compounds can be doubly pluralized, but only when the first noun

is irregular: men-children, menfish, menservants, gentlemen-farmers,

women writers, and women-doctors, but not boys-kings, girlsfriends, or players-coaches. Could the effect have a boring explanation, such as that it sounds funny

to have an -s suffix sandwiched inside a compound?

There is a kind of word

that nicely rules out that possibility. Richard Lederer asks, “Doesn’t it seem

just a little loopy that we can make amends but never just one amend; that

no matter how carefully we comb through the annals of history, we can

never discover just one annal; that we can never pull a shenanigan, be in a

doldrum, or get a jitter, a willy, a delerium tremen, a jimjam, or a heebie-

jeebie?”55 Lederer is alluding to pluralia tantum: Nouns that are always plural. Because they are not the result of pluralizing a singular, the complete

plural form, -s and all, has to be stored in memory. Pluralia tantum in a

sense are irregular regulars, and indeed they are happy to appear inside

compounds: almsgiver (not almgiver), arms race (not arm race), blues

rocker (not blue rocker), clothesbrush, Humanities department, jeans

maker, newsmaker, oddsmaker, painstaking. 56

(Incidentally, do not be distracted by the inconsistent way compounds

are spelled in English: sometimes as one word, as in teethmarks; sometimes

with a hyphen, as in mice-infested; sometimes as two words, as in geese

crossing. The way to recognize a compound is by its composition, such as

being two nouns in a row, and by its stress pattern: Compounds usually

have their main stress on the first syllable, whereas phrases have their main

stress on the second. Compare blackboard with blàck bōard, dárkroom with

dàrk roōm. )

The linguist Paul Kiparsky explained this effect with an influential

theory.57 Words are built in several stages, like a product on an assembly

line. First there is a lexicon of memorized roots, including, as we would

expect, irregular forms. (Kiparsky actually proposed that this box had rules

generating irregulars, as in Chomsky and Halle’s rules-all-the-way-down

theory discussed in chapter 4, but his explanation works the same way if we

assume irregular forms are stored whole.) The lexicon provides the input to

regular derivational morphology, the rules that create complex words out of

simple words and morphemes like learn + -able, dance + -er, and black +

top. The output of this box, a complex word or stem, is then inputted to a

third box, regular inflection, which modifies the word for its syntactic role

in the sentence: past or present, singular or plural. The schematic for

morphology would look something like this: it is formed from rat by a

regular inflectional rule in the third box, too late to feed the compounding

rule in the second box. Hence we get rat-infested but not rats-infested. The word mice, stored as a root in the first box, is available as an input

to the compounding rule in the second box, where it is joined to infested to

yield mice-infested. Rats, however, is not stored as a memorized root in the

first box;

Kiparsky’s explanation is easy to understand and highlights a qualitative

difference between irregular and regular forms: Irregulars are roots and can

be the input to the process of word formation; regulars are the products of

rules and have to be the output of the process of word formation. But we

should not take the theory too literally and believe that words are always

formed in conveyer-belt style or that anything stored in memory can be

inserted into a compound. Remember from chapter 5 that some regular

plurals are stored in memory, even though they don’t have to be. Being

stored, however, can’t be an admission ticket to a compound: cats

presumably is common enough to have a presence in memory, but cats-

infested sounds bad nonetheless. Instead we should interpret Kiparsky’s

model as laying out the logic of word formation: what kinds of words may

snap together in which ways to form bigger words. The restriction is

actually that the kind of word that must be stored in memory—a root—is

the input to complex word formation. Irregular mice is a root, just like a

plain old noun such as duck or rat, and it may be entered into a compound;

cats and rats are not roots but are syntax-ready words, and are the wrong

kind of entities to enter into a compound. There are counterexamples, as there always are in linguistics. 58 For a

while Annie Senghas and I collected as many as we could find and put them

in—what else?—a counterexamples list. Here are some compounds that

contain regular plurals, contrary to everything I have told you so far:

admissions committee

MIT Innovative Structures Program

Boston Antiques Show

morphemes project

Celtics fan

personals ad

chemical weapons attack

publications catalogue

claims applications

ratings data

comics syndicate

records department

cuts package

repeated measures design

enemies list

singles bar

faces lab

skills gap

gimmicks war

top videos list

grades meeting

twins project

injuries report

unemployment benefits cut

landmarks commission

What is going on?

Could we have deluded ourselves by crowing about

the examples that fit a false theory while ignoring the examples that

contradict it?

It seemed unlikely, given the obvious contrast of naturalness

within such nicely matched pairs as mice-infested and rats-infested or

teethmarks and clawsmarks. But the only way to know is to watch people

deal with new examples. Senghas, Kim, and I made up a new questionnaire

with items like these:59

Hordes of rabid rats are swarming out of the Callahan Tunnel since

construction began there. The governor has given a rats-alert

advising people to stay in their homes. My cat Muffin left three dead mice on my doorstep this morning. She’s a pretty good mice-hunter. The senior fraternity brothers just bought an awful contraption for

hazing week. One by one, each pledge will put one of his feet into

the small box while the fraternity president cranks the feet-crusher

tight. At the ski lodge they have a huge central fireplace with a wooden

rail around it that all the people rest their hands on when they get

cold. It’s the best hands-warmer I know. People rated the naturalness of these compounds, which contain plurals,

and rated the same compounds when they contained singulars: rat-alert,

mouse-hunter, foot-crusher, hand-warmer. The questionnaire came in

different versions, one with rats-alert and mice-hunter, the other with mice-

alert and rats-hunter. This ensured that any differences in the plausibility of

the items themselves would cancel out when we compared the average

ratings of regular and irregular plurals. The outcome was clear-cut. People liked compounds with irregular

plurals, such as feet-crusher, significantly better than compounds with

regular plurals, such as hands-crusher. They liked the singulars best of all

— -foot-crusher, hand-crusher—but when forced to consider the plurals,

they liked the irregulars much better. We were relieved, but still had a mystery to solve. Regular plurals never

sound as good inside compounds as irregular plurals do, and they are

usually reduced to their singular form. Yet sometimes they sound good

enough for people to say injuries list and landmarks commission. Why is

rats-hunter bad but injuries list good?

Why are the boosters of one team

called Jets Fans and the boosters of another called Raider Rooters?

60 We had designed the questionnaire to test various explanations, but none

worked. The mystery was solved by the psychologists Maria Alegre and Peter

Gordon, and the solution comes from the first law of language: Strings are

nothing, trees are everything.61 Alegre and Gordon began with a well-

known phenomenon we encountered in chapter 2 when mulling over

mother-in-laws: Words in English sometimes swallow entire phrases, not

just other words. Here are some examples from the linguist Rochelle

Lieber:62

the Charles-and-Di syndrome

a pipe-and-slipper husband

over-the-fence gossip

off-the-rack dresses

God-is-dead theology

a seat-of-the-pants executive

a who’s-the-boss wink

a floor-of-the-birdcage taste

These compounds cannot possibly come off the conveyer-belt model of

word building, because phrases like off the rack and sentences like God is

dead are assembled by the rules of syntax, not the rules of morphology. The

completed phrase has to be routed backward to the word-formation box,

where it can be joined with dress or theology to form the compound:

The good news is that this loop provides a route by which regular

plurals can appear inside compounds. A plural can be created by a rule of

regular inflection, grown into a one-word noun phrase (NP) in the syntax

box, and sent back to the word-formation box to be inserted into a

compound. The result is a recursive tree like this:63

The problem now is that the theory is in danger of saying, Heads I win,

tails you lose. When we find an irregular inside a compound, we call it a

word; when we find a regular, we call it a one-word phrase. Without some

independent way of telling a word-inside-a-compound from a phrase-

inside-a-compound, we have sacrificed the original explanation for the

difference between mice-infested and rats-infested, and the theory becomes

useless. Alegre and Gordon knew, however, that there are ways to tell a

word from a phrase. One is based on tree structure, the other on meaning. What is a red rat eater?

It could be a rat-eater that is red:

Or it could be an eater of red rats:

The second tree is made possible by the recursive loop that allows a phrase,

in this case red rat, to be injected into the compound. What then is a red rats eater?

Alegre and Gordon’s theory states it

would have to be an eater of red rats (as in the second tree), not a red eater

of rats (as in the first tree).

That’s because the rat + s combination has to be

inside a phrase. It was born in the third box, too late to sneak directly into

the compound, and could only have been admitted via the long loop through

the syntax box and then backward to the word-formation box. Once you let

the syntax build a phrase to accommodate the plural, it can put an adjective

like red in the phrase too. So ask yourself: What do you think a red rats eater is?

If you can only

imagine something that eats red rats, you have confirmed the explanation. You don’t have to ask yourself, however; Senghas and I already have asked

a group of adults, and Alegre and Gordon have asked a group of preschool

children. 64 The adults were shown three-word compounds such as torn

receipts envelope, sometimes with a singular noun ( receipt), sometimes

with a plural noun ( receipts).

They were asked to choose from a pair of

descriptions: “the envelope for torn receipts,” or “the torn envelope for

receipts.” The children had to pick the green spiders eater (or the green

spider eater) from a pair of pictures, one with a green monster eating brown

spiders, one with a brown monster eating green spiders. Everyone

interpreted the compounds as Alegre and Gordon predicted: Adults

interpreted a torn receipts envelope as an envelope for torn receipts;

children interpreted a green spiders eater as an eater of green spiders. This leaves us with a final question: Why do people sometimes use the

loop and say enemies list, but sometimes shun it and say rat-infested?

Perhaps there is a subtle difference in meaning, and if Alegre and Gordon’s

theory is on the right track, it should be predictable from the difference in

meaning between a phrase and a word. Words are generic: dog by itself refers to dogginess. A dog hater

needn’t hate any dog in particular; he may never even have met a dog. Phrases, though, are particular: the dog, my dog, and a big dog single out

particular canines.65 Alegre and Gordon noticed that in most of the items in

our counterexamples list, the plural noun referred to a heterogeneous

collection of individuals, each treated as a distinct entity. The whole point

of an enemies list is to keep tabs on particular people, and a publications

catalogue names publications any one of which a reader might want to

order. But when your apartment is rat-infested, one rat is as good or bad as

another, and when apprehending a set of clawmarks, few people ponder

every claw that made a mark. As usual, we need an experiment to elevate the convenient story into a

genuine explanation. Alegre and Gordon asked one group of subjects to rate

the heterogeneity of the referents of the first noun in each of a set of

compounds, and asked another group to rate the naturalness of the

compounds when the first noun was plural. The heterogeneous nouns, as

they predicted, made better compounds-with-plurals-inside. In science the pursuit of idle curiosity often pays off in deeper

understanding. People’s inquisitiveness about flied out, talismans,

sabertooths, still lifes, outputted, rat-infested, and other unexpected forms

has enhanced our understanding of regularity and irregularity and provided

an entirely new kind of evidence in favor of the words-and-rules theory. The examples themselves show that regular and irregular forms are

qualitatively different, not merely endpoints on a continuum of

predictability. The regular pattern can apply to special words such as fly out

and still life; the irregular patterns cannot. Irregular plurals can easily go

inside compounds such as teethmarks and mice-infested; regular plurals

cannot. The examples also show that people consider more than sound when

they form new words: An input sound like fly can emerge in the output as

either flew or flied, depending on the person’s analysis of the whole word. But people consider more than meaning, too; a word’s meaning may be

stretched to a wispy filament, as in threw up or cut a deal, and people will

inflect it as if nothing had happened. Instead, people are instinctive linguists, assigning a structure to every

word. We all tacitly judge whether a word is a canonical root or some other

kind of sound, and we analyze how the word may have been constructed

from other words. The analyses that lie behind people’s choice of regular

and irregular forms have deepened our understanding of the nature of words

and rules. The most basic kind of word is a root, with a canonical sound

arbitrarily paired with a meaning and a part of speech. The most basic

action of rules is to compute the properties of a complex form from the

properties of its parts and way they are arranged, with a special role for one

position in the arrangement, the head. Like many psycholinguists I have always thought of language in terms

of rules and structures, but I was never sure I could defend the attitude

against a hardened skeptic. It was only when I learned about the phenomena

in this chapter that I became convinced that rules are living things

inhabiting people’s minds. The theory that rules are a major ingredient of

language offers a reasonable enough explanation for how we inflect new

words, rare words, and unusual words. But when the theory also gives a

click of insight into why we say lowlifes instead of lowlives and why teeth

leave teethmarks but claws leave clawmarks, and ten other mysteries, it

begins to take on the ring of truth. All the more so when it survives a wave

of assaults by dangerous counterexamples. The phenomena we have examined also provide us with a set of

instruments to probe people’s mental representation of words when we turn

in the coming chapters to three great challenges for the words-and-rules

theory. Can we catch children in the act of learning a rule as they master

their mother tongue?

Do rules work in all the world’s languages the way

they do in English?

And can we distinguish words and rules in the human

brain?

7

KIDS SAY THE DARNEDEST THINGS

Moving its translucent mass through the watery

shadows of the dock and then, past the dock (some-

thing so real which now is not), the jellyfish

swam in its slow float while we (I and my daughter,

then just three) ran back and forth predicting that limp

pink gleam and each embodiment it would seem. “A jello umbrello!” she began and turned

to me expectantly. Censoring (an after-

birth, broken veins, or Medusa’s myth, the monstrous

queen made mortal and mother), I stood in silence

until it ended with a shout: the jelly-

fish glided out. Now months have passed, but surprise!

“The jellyfish was in my eyes!” Caroline calls

while caught between depth and surface of a dream. “It bleeded and it singed!” Her conjugations

soon will exact simple irregularities

and tensing will be not verbs, but time’s tentacles

untangling her parachute, waving at me.1

—Susan Kinsolving, “The Jellyfish,” 1999

Grammatical errors like bleeded and singed have long epitomized the

innocence and freshness of children’s minds. The errors are acts of creation,

in which children lift a pattern from their brief experience and apply it with

impeccable logic to new words, unaware that the adult world treats them as

arbitrary exceptions. In A Dark-Adapted Eye, the novelist Barbara Vine

introduces an unlikable child by remarking, “He would refer to ‘adults’

instead of ‘grownups,’ for instance, and get all his past tenses right, never

saying ‘rided’ for ‘rode’ or ‘eated’ for ‘ate.’”2

Children’s errors with irregular verbs also have been prominent in

debates on the nature of language and mind. The neurologist Eric

Lenneberg pointed to the errors when he and Noam Chomsky first argued

that language was innate; the psychologists David Rumelhart and James

McClelland set them as a benchmark when they first argued that language

could be acquired by generic neural networks. Psychology textbooks cite

the errors to rhapsodize that children are lovers of cognitive tidiness and

simplicity; researchers who study learning in adults cite the errors as a

paradigm case of the human habit of overgeneralizing rules to exceptional

cases. 3

Nothing is more important to the theory of words and rules than an

explanation of how children acquire rules and apply them—indeed

overapply them—to words. The simplicity of these errors is deceptive. As

we shall see, it is not easy to explain why children start making them, and

it’s even harder to explain why they stop. Overgeneralization errors are a symptom of the open-ended productivity of

language, which children indulge in as soon as they begin to put words

together. At around eighteen months children start to utter two-word

microsentences like See baby and More cereal. 4 Some are simply telegraphic renditions of their parents’ speech, but many are original

productions. “More outside!” says a tot who wants to play in the park. “Allgone sticky!” says another after his mother has washed jam off his

fingers. My favorites in the data from my own lab are “Small loud” after

someone had turned down the stereo, and “Circle toast!” shouted repeatedly

to uncomprehending parents who couldn’t figure out that the child wanted a

bagel. 5

By their twos, children produce longer and more complicated sentences,

and begin to supply grammatical morphemes such as -ing, -ed, -s, and the

auxiliaries. 6 Sometime between the end of the second year and the end of

the third year, children begin to overgeneralize -ed to irregular verbs. All

children do it, though parents don’t always notice it. My sister told me that

her son Carl never made this kind of error, and as if to contradict her, he

said sticked in my presence a minute later. When children are old enough to

sit still in experiments, they pass the wug-test: After hearing that a man

knows how to rick or bing, they say that yesterday he ricked or binged. 7

Children regularize almost anything they can. They put -ed not only on

irregular stems, as in breaked and eated, but on irregular past-tense forms,

as in broked and ated. They put it on their own neologisms, such as

poonked, light-ninged, and spidered. They put it on past-tense forms that

already have a suffix, as in sweepened, presseded, and My brother got sick

and pukeded. 8

The past tense is not the only source of irregularity in English, and it is

not the only regular pattern children overgeneralize. Alongside past tense

errors such as breaked and putted we find plural errors such as mans, foots,

tooths, and mouses. 9 Three English verbs are visibly irregular in the third-person singular present tense, and children overgeneralize -s to all three:

He just haves a cold. She do’s what her mother tells her. No, she be’s bad, then she be’s good, OK?10

The suffixes -er and -est turn many adjectives into comparative or

superlative forms. It’s easy to forget that the rule has exceptions until we

hear children adding suffixes to them. They overgeneralize the suffixes to

polysyllabic adjectives, as in specialer and powerfullest, and to a handful of

suppletive irregulars:11

“Wow!

I must’ve been gooder than I thought!”

THE FAMILY CIRCUS. Reprinted with special permission of King Features Syndicate. Children often generalize from fourth, fifth, and sixth to oneth, twoth, and threeth, or sometimes firstth, secondth, and thirdth. They leap from myself, yourself, and herself to hisself, and from ourselves and yourselves to theirselves. I have heard of one child who used its rather than them as the

plural of the pronoun it, and another who liked drawing rectangles,

triangles, and cirtangles (circles).

12

Children are overzealous grammarians not only in applying inflections

in their own speech but also in analyzing them in the speech of others. They

have little choice. Children are never given grammar lessons presenting -ed

or -s with lists of stems to conjugate or decline; they must mentally snip the

suffixes out of the full, inflected words they hear in conversation. As they

are figuring it out, they occasionally snip too eagerly and come out with

strange back-formations:

FOR BETTER OR WORSE reprinted by permission of United Feature Syndicate, Inc. I suspect that comic strips showing a child making a speech error are

usually based on real-life instances known to the cartoonist; in almost every

example I have seen, similar errors have been documented in the scientific

literature. Alan Prince studied a girl who, like April, was delighted by her

discovery that eats and cats were really eat + -s and cat + -s. She used her new suffix snipper to derive mik (mix), upstair, downstair, clo (clothes), len

(lens), sentent (sentence), bok (box), brefek (from brefeks, her word for breakfast), trappy (trapeze), even Santa Claw. 13 Another child, overhearing his mother say they had booze in the house, asked what a “boo” was. One

seven-year-old said of a sports match, “I don’t care who they’re going to

verse,” from expressions like the Red Sox versus the Yankees. 14

We laugh, but adults do the same thing, or at least our ancestors did. Cherry is a back-formation from cerise, and pea is the invented singular of

the mass noun pease, as in the nursery rhyme “Pease porridge hot, pease

porridge cold.” (Perhaps someday a grain of rice will be known as a rouse. )

Many people have to be reminded that there is no such thing as a kudo: The

noun kudos is singular, from the Greek word for glory. A striking feature of children’s past-tense errors is that they appear,

sometimes suddenly, after long stretches in which the children use the past

tense correctly when they use it at all. A child might say sang, went, and

heard for many months before coming out with singed, goed, and heared. 15

In a sense, the child gets worse before getting better; if the percentage of

past-tense forms of irregular verbs that are correct is plotted over time, the

shape of the graph looks a bit like a U. “U-shaped development” fascinates

child psychologists because with almost anything else you measure,

children get better as they get older. 16 No one considers childhood to be a

period of decline (that comes later), so the newly appearing errors are taken

as a sign of a reorganization in the child’s mind. A laundry list of

disconnected items suddenly reveals itself as having a pattern, and the child

extracts the pattern and applies it across the board. In the case of the past tense, children have a smattering of regular forms

such as played and used before they make their first error with an irregular,

and they use them correctly to talk about events in the past. 17 Presumably

they have memorized those forms as indigestible chunks and use them like

any other word, with the “pastness” simply being part of their meanings. At a certain point a child notices (not consciously, of course) that many

words come in ever-so-slightly different versions: walk and walked, use and

used, play and played, push and pushed. Logically speaking, these could be

interpreted as meaningless variations in pronunciation or speaking style, yet

something impels children to seek a principle behind the variation. By

subtracting walk from walked, push from pushed, and so on, a child can isolate -ed. By correlating its use with its meaning—that is, noticing that

Mom and Dad use -ed when describing events that are over and done with

—the child can infer that -ed means “past tense.” This synopsis brushes

aside many complexities, such as how the child knows to look out for

“present—past” instead of “hot—cold,” “indoor—outdoor,” “good mood—

bad mood,” and hundreds of other interesting distinctions. It also sweeps

aside how a child deduces that the rule is obligatory: You can’t say I

already eat breakfast this morning, even though the meaning would be

clear. Yet children do succeed, and once the rule has been discovered they

can feed any verb into it, regular or irregular. They now can say goed and

heared and bleeded and singed in situations where earlier they might have

said went and heard and bled and sang. 18

Unfortunately, the rule-epiphany theory by itself cannot explain why

children make errors like bleeded and singed. I have said that children start

saying bleeded and singed because they have acquired an “Add -ed” rule. But adults have an “Add -ed” rule too, and we don’t say singed. (If we did,

we wouldn’t call the child’s form an error.) Something important is missing:

the difference between children and adults, and how children overcome the

difference as they grow up. A first guess is that children become adults because language

development is driven by communication: Children improve their language

in directions that allow them to communicate their wishes more effectively. Wrong. There is nothing unclear about the meaning of bleeded or singed. In

fact, as long as children make these errors, their language is more

communicative than adults’.

English has about twenty-five irregular verbs

that don’t change their forms in the past tense, such as cut, set, and put. These verbs are ambiguous between past and nonpast: On Tuesday I put the

trash out could mean last Tuesday, next Tuesday, or every Tuesday. The

childlike form On Tuesday I putted the trash out could mean only a

preceding Tuesday. A language is certainly a powerful tool for

communication, but children could not acquire its details by figuring out

which ones help in communication; they learn the whole language, with all

its strengths and weaknesses, because they just can’t help it. A second guess is that we adults don’t say bleeded and singed because

we never hear other adults say them. Wrong again. Adults say lots of things

they never hear other adults say. New verbs constantly enter the language—

to diss, to snarf, to fax, to mung, to wild, to flame, to mosh—and an adult

who learns diss in the present tense does not have to wait to hear someone

say dissed before using it in the past tense. If adults say dissed even though

they have never heard it, they should be willing to say singed even though

they have never heard it. The reason adults avoid making regularization errors is not that the error

has never been heard; it’s that the irregular counterpart has been heard. There must be a component of adult psychology that causes the experience

of hearing an irregular form such as sang to inhibit the application of the -

ed rule to that item. As noted in chapter 5, this component is called blocking: A specific form in the mental lexicon blocks the application of a

general rule that would express the same grammatical notion (past tense,

plural, and so on), perhaps through an inhibitory link from the lexicon to the

rule.19 Thus sang, listed as a past tense of sing, blocks the past tense rule, preempting singed; geese, listed as the plural of goose, blocks gooses;

better, listed as a comparative of good, blocks gooder. Perhaps, then, children lack the blocking principle and have to learn it. But how?

To learn the blocking principle children first would have to know

tht forms like singed are ungrammatical. Remember, not hearing other

people say singed isn’t enough, because other people don’t say wugged

either, and may not say munged or flamed, but people do not avoid the

unheard past-tense forms. The only way for children to know that singed is ungrammatical is to

use it and get a negative feedback signal from their parents—a correction, a

frown, a puzzled look, or a non sequitur as a response. Information about

what is not in a language is called negative evidence, and it is one solution

to what linguists call “the logical problem of language acquisition”: how a

child could, in principle, learn an entire infinite language from a finite

sample of the behavior of its speakers.20

Children almost certainly do not solve the language acquisition problem

by depending on negative feedback from parents. For one thing, parents

could not very well correct or disapprove of their young children every time

they err. Most of toddlers’ sentences are ungrammatical in some way, so

parents would be chiding them all day long. Parents focus on the content of

their children’s sentences, not their form, and let most errors slip by:

FATHER: Where is that big piece of paper I gave you yesterday?

ABE: Remember?

I writed on it. FATHER: Oh that’s right don’t you have any paper down here buddy?

21

What happens when parents do correct their children?

The cartoonist

Bill Keane shows two of the results:

“Mommy, Dolly hitted me.”

“Dolly HIT me.”

“You too?!

Boy, she’s in trouble!”

THE FAMILY CIRCUS Bil Keane, Inc. Dist. By Cowles Synd., Inc. THE FAMILY CIRCUS Bil Keane, Inc. Dist. By Coudes Synd., Inc. Keane has a fine ear for children’s language, and the dialogues are not

fanciful. Here is a real one, transcribed by the psychologist Courtney

Cazden:22

CHILD: My teacher holded the baby rabbits and we patted them. ADULT: Did you say your teacher held the baby rabbits?

CHILD: Yes. ADULT: What did you say she did?

CHILD: She holded the baby rabbits and we patted them. ADULT: Did you say she held them tightly?

CHILD: No, she holded them loosely. Systematic studies bear out the anecdotes. The linguist Arnold Zwicky,

observing his daughter’s overgeneralization of participle endings, reported

that “six subsequent months of frequent corrections by her parents had no

noticeable effect.” 23 The psychologists James Morgan and Lisa Travis

looked at transcribed speech of three children and their parents, sampled

biweekly for several years. They wanted to see whether the children’s errors

elicited any consistent pattern from their parents—not only overt

corrections, but partial or full repetitions, requests for clarification,

questions, attempts to move the conversation on, or silence. No consistent

pattern was found. In a follow-up study, Morgan and Travis found a

different kind of grammatical error, in which parents do sometimes recast a

child’s sentence in correct English. But they found that the recasting had no

effect—if anything, it had an adverse effect—on the child’s subsequent

improvement.24

The psychologist Karin Stromswold has a particularly dramatic

demonstration that parental feedback cannot be crucial. She studied a child

who, for unknown neurological reasons, was unable to talk, but who was an

avid listener and understood complex sentences. When the boy was four,

Stromswold tested his knowledge of past-tense forms by asking the boy to

teach a dog puppet to talk. She asked him to give the dog a bone when it

spoke correctly and a rock when it made an error. The boy awarded bones

for heated, baked, showed, and sewed, and rocks for eated, taked, and knowed. He made just one error, awarding a bone for goed, similar to the

performance of normal children. Somehow the boy, and presumably other

children, can come to recognize that overgeneralized forms are

ungrammatical without first having to make the errors and note their

parents’ response. 25

Children must solve the logical problem of language acquisition in a

different way. Perhaps, rather than learning the blocking principle from

evidence that singed is not English, they begin with the blocking principle

and use it to deduce that singed is not English. That is, blocking might be

built in to the circuitry that drives language acquisition—what Chomsky

calls Universal Grammar and what I have called the language instinct. As

with all sane proposals about innate structure, such an instinct would not be

an alternative to learning but rather an explanation of how learning works. In this case, because children hear parents say sang in the course of

ordinary conversation, they retain sang in memory, and the blocking

mechanism represses their tendency to say singed, turning them into

adults.26

We need one more assumption to get the theory to work. If children

already have blocking, and all else is the same, they should never say

singed to begin with!

Having heard their parents say sang even once should

be enough to block the rule from applying to it. Fortunately, the extra

assumption is as parsimonious as a theory in child psychology can be. What is the simplest conceivable hypothesis of how children differ from

adults?

Answer: They have not lived as long. That is what being a child

means. Now, among the experiences we accumulate as we live our lives is

hearing the past-tense forms of irregular verbs. Human memory profits

from repetition. If children have heard sang less often than adults have,

their memory trace for it will be weaker and their ability to retrieve it will

be less reliable. Sometimes, when they are trying to express the thought

“singing in the past,” sang will not pop into mind (or at least not quickly

enough to get put into the sentence).

Before children acquire the -ed rule,

when they fail to retrieve sang they have no choice but to use the bare stem

sing, even for events that happened in the past. But once they have acquired

the rule, they can apply it to sing, creating singed, thereby satisfying the syntactic constraint that tense be marked in every sentence. This minimalist theory combines a simple idea from linguistics

(blocking) with a simple idea from psychology (memory improves with

repetition).

It explains why children get worse before they get better, and

solves the logical problem of how they exorcise their errors without

parental feedback. Correct forms such as sang that a child used early on do

not go anywhere once the child has acquired the rule, nor are they incapable

of blocking errors: They simply must be retrieved from memory to do the

blocking, and they are not always retrieved. The cure for overgeneralization

is living longer, hearing irregulars more often, and consolidating them in

memory, improving their retrievability. Indeed this account, which posits that the mind of a child and the mind

of an adult work the same way, is deducible from the very logic of

irregularity, augmented only by the fact that memory is fallible. What is the

past-tense form of the verb to shend, meaning “to shame”?

If you answered

shended then you have overgeneralized; the correct form is shent. This

“error,” of course, is to be expected. Irregular forms, by definition, are not

predictable, so the only way you could have produced shent is if you had

previously heard and remembered it. But you have heard it zero times and

can’t have remembered it. If in two years you were asked the question and

erred once more, it still would not be surprising, because you would have

heard it only once. Now put yourself in the child’s shoes. Many verbs will

be like shent for you: never heard, or not heard enough times to be

recallable on demand. The mystery of why children say singed and bleeded

has been solved. When children say singed, are they simply little adults with bad memories?

Gary Marcus and I combed through computer files with transcripts of the

spontaneous speech of 83 children and extracted 11,500 sentences with

irregular past-tense forms. We wanted to figure out when and why children

start making errors, how often they do it, and with which verbs. Most of

what we found fit the simple theory. 27

First we looked at the error rate. If a child’s language system is basically

like an adult’s, it should be designed to suppress the regularization of verbs

that the child remembers are irregular. The suppression cannot be perfect

because memory is not perfect, but children’s memory for words ought to

be fairly good; the child is, after all, using thousands of words and acquiring

a new one, on average, every two hours. Overgeneralization errors should

be the exception, not the rule, coming from the occasional breakdown of a

system built to prevent the error. In fact the average error rate across

children is only 4 percent. More than 95 percent of the time a child utters

the past-tense form of an irregular verb it is a correct form like sang, not an

error like singed. (Adults tend to overestimate the error rate because they

remember the errors, which stick out like sore thumbs, and fail to notice the

boring correct forms.) Once children begin to make the errors in their third

year, they continue at this low rate until well into the school-age years. No verb is immune to the errors, not even those a child used correctly

before the error-making began. Nor is any verb consistently erred on. A

child might us e felt when young, then both felt and feeled when somewhat

older. The errors are haphazard; children sometimes use correct and

incorrect versions in quick succession, like this: “Daddy comed and said

‘hey, what are you doing laying down?’ And then a doctor came....” 28

The hit-or-miss nature of the errors suggests that children are not ignorant

of the correct forms; they are fallible at retrieving them. Some verbs are

more error-prone than others, and the simple theory predicts that these

should be the verbs that the child has heard less often. So we counted how

often the children’s parents used each irregular verb in the past tense. If a

parent used told and brought more often than, say, froze and won, the child should have a stronger memory trace for told and brought than for froze and

won, and should say telled and bringed less often than freezed and winned. We examined ninety irregular verbs and found that with every child, the

more often the child’s parents used a verb in the past tense, the less often

the child regularized it. Could children on some level really know that their errors are errors?

Sometimes they do. The psycholinguists Dan Slobin and Tom Bever tried

using their children’s errors in their own speech, just for fun.29 The

children were not amused:

TOM: Where’s Mommy?

CHILD: Mommy goed to the store. TOM: Mommy goed to the store?

CHILD: NO!

( annoyed) Daddy, I say it that way, not you. CHILD: You readed some of it too...

she readed all the rest. DAN: She read the whole thing to you, huh?

CHILD: Nu-uh, you read some. DAN: Oh, that’s right, yeah. I readed the beginning of it. CHILD: Readed?

( annoyed surprise) Read!

( pronounced red)

DAN: Oh, yeah, read. CHILD: Will you stop that, Papa?

In more controlled studies children are asked to judge the past-tense

forms of a language-impaired puppet. They let many errors slip by, but they

object to errors more often than to correct forms. And when asked to

choose, the children, on average, prefer the correct forms. 30 All this

suggests that children really do know irregular past-tense forms like went

and read; their errors must be slip-ups in which they cannot slot an irregular

form into a sentence in real time. If overgeneralizing children are not qualitatively different from adults,

we should see adults making the errors, and indeed they do, approximately

once in every 25,000 sentences in which they use an irregular past-tense

form. 31 This figure is about a thousand times less frequent than children’s

errors, but the estimate includes common verbs like came and went and told

that have been drilled into our heads tens or hundreds of thousands of times. With the less common irregulars adults make “errors” quite often. It’s hard

to say how often, because we adults get to say what counts as “correct,” and

if we regularize an irregular often enough, we simply declare by fiat that it

is not an error!

These muzzy alternatives— dreamed and dreamt, pleaded

and pled, leaped and leapt, strided and strode—are lower in frequency than

pure irregular verbs like went and came, much as children’s errors such as

winned tend to occur with the verbs they hear less often. Even among pure

irregular verbs, those used with lower frequency like slew and strove are

judged to be somewhat unnatural. and their regular counterparts are judged

to be relatively unobjectionable.32

Over the long run this psychology changes the composition of a

language. Say you have heard strode only a few times in your life—more

often than shent, but far less often than held. You would have a weak

memory trace for strode, just strong enough for you to recognize it and for a

little voice in your mind’s ear to whisper, “strode!”, but not strong enough

to block the regular rule from applying. You may very well say strided, just

as a child would say hided. If your neighbors are similarly ambivalent, the

language community may be divided, with some people saying strided,

others saying strode, and still others, hearing their neighbors using both

forms without rhyme or reason, memorizing both and using them

interchangeably. With rarer verbs adults’ “errors” create a vicious circle: They use an

irregular form less and less, so their children and neighbors hear it slightly

less often, causing their memory traces in turn to be weaker, causing them

to use it less (and regularize it more), in turn causing their children and

neighbors to hear it less, and so on. An irregular form that falls below a

critical frequency could disappear outright after a few generations. As we

saw in chapter 3, that is exactly what has occurred in the history of English:

The irregular forms of less common verbs such as chide—chid, cleave—

clove, and geld—gelt became extinct.33 Verbs, like all bits of culture, can rise or fall in popularity, and one can imagine a time when the verb to geld

had slipped so far that a majority of adults lived their lives without having

heard gelt. When pressed, they would have used gelded; the verb had

become regular for them, and for all subsequent generations. That is why

irregular verbs tend to be high in frequency; the list has been filtered

repeatedly through the minds of children and adults, both of whom

regularize uncommon irregular verbs. What launches the transformation from regurgitating correct forms to

creating incorrect ones?

Why does a child wake up one morning and start to

say bleeded and singed?

The simplest theory is that that is precisely the point at which the child

has acquired the past tense rule, a result of the process described on page

193.

The rule must be acquired at some point; it could not possibly be

innate, because some languages don’t mark tense on their verbs, and those

that do don’t use the English -ed. Prior to learning the rule, a child with an

irregular form stuck on the tip of her tongue could do no better than to utter

the bare stem sing; with the rule in hand she can fill the vacuum with

singed. One way to confirm the theory is to watch what happens to regular

verbs when the child makes the first error with an irregular. Before the first

error children leave regular verbs unmarked most of the time; they say

Yesterday we walk. Then they begin to mark these verbs most of the time, as

in Yesterday we walked. It is during this transition that the first error with an

irregular form, like singed or heared, appears. We can interpret the tandem

development of walked and singed as two signs of a single underlying

process, the acquisition of the -ed rule: correct performance where the rule

is called for, and errors where it is not. 34

This idea that children add a rule onto their list of words is simpler than

the suggestion that children radically reorganize their language, abandoning

the list in favor of an imperialistic rule system and then slowly reacquiring

the list. The simpler idea also fits the facts better. When children begin to

make these errors, where do the four percentage points of errors come

from?

Are they produced in situations where previously the child would

have produced a correct form such as sang?

Or are they produced in

situations where the child would have produced a bare stem like sing?

That

is, do the errors drive out correct forms, a mysterious step backward on the

road to adult language, or is one kind of error driving out a different kind?

The data say that the errors of commission ( singed) are driving out other

errors, errors of omission ( sing), not correct forms ( sang).

For example,

before making his first overt error, one boy we studied used correct forms

74 percent of the time and produced the bare verb 26 percent of the time. When the errors began, at a rate of 2 percent, did they come out of the 74

percent of his verb usages that were correct, driving performance down to

72 percent?

No—the correct forms increased, to 89 percent; the two

percentage points of new errors came at the expense of the errors of

omission, which dropped to 9 percent. Children don’t backslide; when they

supplement sang with singed, they take a step forward, because the syntax

of the sentence, which demands a past-tense form, is satisfied more often.35

What triggers the “Eureka!” moment, when a child first discovers a

rule?

Why does it dawn on some children in their late ones, but on others

not until their late twos?

I suspect we will never understand what triggers

the very first error. Two children we studied made no errors for seven or

eight months, popped out a single error ( feeled or heared) just before

turning three, and then went another five months before doing it again. Why

the false start?

What were the children thinking in the months when they

failed to act on their epiphany?

One possibility is that the gap is an illusion

of sampling. Perhaps a newborn rule is wobbly and unreliable, and there are

only so many times a child has the urge to use an irregular verb in the past

tense, fails to retrieve its stored form, runs the rule to completion, and the

tape recorder is running. A steady low probability in the mind of the child

may surface as sputtering fits and starts in a record of the child’s speech. Another possibility is that language development at times really is

chaotic, because the child is trying to make sense of the language with a

changing brain. Synapses, the connections between brain cells, sprout and

die in large numbers in the first few years of life, and the churning may

temporarily swamp or wash away the newly laid down trace of a rule. Also,

countless random events affect the microscopic structure of the growing

brain. The human genome does not have nearly enough information to

specify the wiring of the brain down to the last connection. We see this in

identical twins, who share all their genes and most of their experiences, and

who have similar, but not identical, brains, intellects, and personalities. 36

Jennifer Ganger and I suspected that at least some of the timing of

language development, including the past tense rule, is controlled by a

maturational clock. Children may begin to acquire a rule at a certain age for

the same reason they grow hair or teeth or breasts at certain ages. If the

clock is partly under the control of the genes, then identical twins should

develop language in tighter synchrony than fraternal twins, who share only

half their genes. We have enlisted the help of hundreds of mothers of twins

who send us daily lists of their children’s new words and word

combinations. The checklists show that vocabulary growth, the first word

combinations, and the rate of making past-tense errors are all in tighter

lockstep in identical twins than in fraternal twins. The results tell us that at

least some of the mental events that make a child say singed are heritable. The very first past-tense error, though, is not. When one twin makes an

error like singed for the first time, an identical twin is no quicker to follow

suit than a fraternal twin. These gaps—an average of thirty-four days

between the first past-tense errors of two children with the same genes

exposed to the same speech—are a reminder of the importance of sheer

chance in children’s development. 37

I have explained children’s creative errors by crediting them with a rule, but

there is an alternative: Children might analogize from words they already

know. They might say holded because hold sounds like fold, mold, and scold, whose past-tense forms are folded, molded, and scolded. Even with

verbs like sing and ring, which do not rhyme exactly with any common

regular verb, children could be reminded of bits and pieces of similar verbs

like sipped, banged, rimmed, and rigged, and cobble together analogous

singed and ringed. That of course is the basis of the pattern associator memories developed

by Rumelhart, McClelland, and their connectionist followers. Rumelhart

and McClelland’s model acquired hundreds of regular and irregular verbs,

generalized to dozens of new verbs, and strikingly, appeared to go through a

U-shaped sequence, first producing correct past-tense forms for irregular

verbs and later overgeneralizing -ed to them. Yet the model had nothing that

looks like a word, a rule, or a distinction between regular and irregular

systems. How did they get a memory model to learn in a way that everyone

has always taken to be a hallmark of rules?

38

They had an ingenious idea. Rumelhart and McClelland figured that

children acquire common verbs first, rarer verbs later. Since common verbs

tend to be irregular, and rare verbs regular, the mixture of irregular and

regular verbs in children’s vocabularies should shift toward the regulars as

their vocabulary grows and they begin to run out of irregulars and

encounter more and more regulars. Moreover, children’s vocabulary growth

shows a big spurt several months after they learn their first words. That

spurt could cause a sudden influx of regular verbs. Pattern associator memories are highly sensitive to changes in the

statistics of their input. When given a small number of oddball items, they

memorize their patterns individually; when given a torrent of items sharing

a pattern, they go with the numbers, extract the pattern, and run roughshod

over the individual items, gradually reacquiring them over many subsequent

bouts of training. That sounds a lot like children. Rumelhart and McClelland imagined an extreme case: A child first

learns a few common verbs, mostly irregular, followed by a spurt of

hundreds of verbs, mostly regular. They consulted a list of word frequencies

in English, found the dividing line that gave the strongest contrast in the

mixture of regulars and irregulars, and trained their network in two stages

accordingly. First the model was fed the 10 most frequent verbs in the list,

80 percent of which were irregular, 10 times apiece; then it was fed the next

410 verbs, 80 percent of which were regular, 190 times apiece. The

network learned the first ten easily. Then, when bombarded by regular

verbs, it strengthened thousands of connections to -ed, which overwhelmed

the connections to the irregular pasts and led the model to make errors such

as breaked. Connectionist modelers following in their footsteps used more

sophisticated networks, but they also induced childlike behavior by

changing the models’ diet of regular and irregular verbs over time.39

Did the computer really “mimic the brain,” as the headlines put it?

It all

depends on whether children begin to say breaked in response to an influx

of regular verbs. Michelle Hollander and I checked the transcripts of the

speech of three children over several years to see whether parents at some

point start using more regular verbs when talking to their children. They did

not. The proportion of regular verbs in parents’ speech—about 25 to 30

percent—is the same when their children are two as when they are five. At

first that may seem odd: when children are young, parents should favor the

common, irregular verbs such as make and do; only when their children are

older should they dip into the lower frequencies and use regular verbs such

as abate, abbreviate, and abhor. The reason that scenario doesn’t appear in

the statistics is that common irregulars like make, do, and hold are

indispensable, general-purpose verbs that people of all ages depend on in

every conversation. Abate, abbreviate, and so on compete with one another

for air time, so even when the number of different regular verbs rotating in

and out of conversation increases, the proportion of conversation filled with

regular verbs remains constant. Perhaps then we should be looking not at the number of times the verbs

are used but rather at the number of verbs in the child’s vocabulary, each

counted once. There the proportion of regular verbs must increase, because

there are only so many irregular verbs in the language, and when they begin

to run out, the child has to acquire more and more regulars. That’s how

Rumelhart and McClelland derived their prediction. But counting

vocabulary items is a bit odd if you think about the actual events that make

up language acquisition. Children presumably learn as they listen to the

speech coming out of the mouths of their parents, not by scanning their own

mental dictionaries and feeding each verb into their network once per pass. We wanted to be charitable, though, so we checked the transcripts to see

whether there really is a vocabulary spurt, and thus a richer mixture of

regular verbs, when children begin to overapply -ed. There was not. Children’s vocabularies spurt in the mid-to-late ones,

about a year too soon to trigger their past-tense errors, which begin in the

mid-to-late twos. In the years in which children make the errors, regulars

are coming in at a lower rate than they were earlier, when the children were

using the irregulars correctly. The timing is not what we would need to get a

pattern associator to overgeneralize after an early stage of correct

performance. The general problem is that Rumelhart and McClelland balanced their

model on a knife-edge of assumptions about the statistics of the speech

input to the child. But language acquisition is a robust process that does not

live or die by the nuances of parental speech statistics. Throughout the

world’s cultures, children must learn the combinatorial tools of their

language across a wide range of input mixtures, as we will see in the next

chapter. Closer to home, even the English plural shows statistics unlike

those of the past tense. The handful of irregular nouns known to children

( men, children, feet, teeth) never could dominate their noun vocabularies

the way that irregular verbs, at least in theory, could dominate their verb

vocabularies. Yet children show the same U-shaped development with

plurals as with past-tense forms. When they begin to speak, all of their

plurals are correct, and then they begin to overgeneralize at a low rate for

several years.40

Michael Ullman and I gave the pattern associator model two more

chances to prove that it mimics the human brain. If children, like the model,

learn by analogy, their irregular verbs should be lured into error by similar-

sounding regular verbs and protected from error by similar-sounding

irregular verbs. If holded is an analogy from soundalik e folded, then the

more soundalike regulars a verb has, and the more frequent they are, the

more likely the verb will be regularized. Holded might be more common

than singed, for example, because holded is strongly attracted to frequent

folded and to a lesser extent scolded and molded, whereas singed is weakly

attracted to low-frequency blinked and not much else. But when we

correlated the number of potential seducers of a verb with its error rate in

children’s speech, we found little to no effect. 41

The model did mimic the brain in one way. If drank owes its survival to

similar irregular forms in memory such as sank and rang, then verbs with

more irregular allies, and more common ones, should be erred on less often. And indeed they are. This difference—irregular forms need similar

irregulars, but regular forms do not need similar regulars—parallels the

findings from adults discussed in earlier chapters. It reinforces the

compromise conclusion that pattern associators capture something about

irregular forms and the memory in which they are stored, but fail to capture

the nature of regular forms and the system in which they are computed. What have children actually acquired when we say they have acquired a

past tense rule?

Is it just one more noise they can make, or is it the powerful

combinatorial tool that, in conjunction with the rest of grammar, gives rise

to the vast expressive range of a language and the elegant logic behind its

apparent quirks?

Children’s past-tense and plural rules really do seem like wobbly

versions of the adult’s, with their sweeping power to inflect any verb or

noun. Children apply their past tense rule to almost all their irregular verbs,

despite the strong associations to irregular past-tense forms. They apply it

to unusual-sounding verbs of their own creation, such as lightninged,

smunched, and poonked. They apply the rules to words built out of phrases,

such as eat lunched, cut-upped egg, and There is two Empire Strikes Backs. Bilingual children sometimes apply a rule to words from their other

language, as in perachs and sefers, Hebrew for flower and book. 42

Children also apply the rule to rootless and headless words, the ones

that lead to such curiosities as lowlifes and flied out, explained in the

preceding chapter. Kim, Marcus, and I gave children a wug-test with a

twist. Half the new verbs were identical in sound to irregulars but obviously

were based on nouns, like to fly meaning “to cover a piece of paper with

flies” and to ring meaning “to put a ring on something.” These are precisely

the circumstances that for adults turn irregular-sounding verbs into regulars

— flied out to right field; high-sticked the goalie; ringed the city with

artillery—because a verb based on a noun lacks a root or head and cannot

tap into the system of irregular roots stored in memory. Children as young

as four work the same way. They regularize verbs based on nouns (as in She

flied the paper) more often than they regularize verbs with verb roots (as in

They are flying down the road).

43

In a similar experiment children saw objects labeled with irregular

nouns. Some were simple noun roots, such as a fuzzy mouse and a little

goose; some were based on names, such as a Mickey Mouse and a Mother

Goose; and some were bahuvrihi compounds, such as a snaggletooth (a

walruslike creature) and a higfoot. When asked to describe collections of

these toys, the children used regular plurals for names and headless

compounds ( Mickey Mouses, Mother Gooses, snaggletooths, bigfoots) more

often than for the simple noun roots ( fuzzy mouses, little gooses).

Children,

like adults, don’t just listen to a word’s sound when they compute its

inflection; they also analyze its grammatical structure. Children are also sensitive to the other curiosity of irregular nouns

discussed in the preceding chapter: the contrast between mice-infested,

where an irregular behaves like any other word and can be inserted into a

compound, and rat-infested, where the regular plural rats is computed too

late to be inserted into a compound. I know of one child who insisted to his

father that a building with mice was a mice building, and another child who

said, “These aren’t only handcuffs; they can be feetcuffs.” I have never

heard of a child say anything like rats building or handscuffs, though of

course we need an experiment to show the difference is real. Peter Gordon introduced a set of three- to five-year-olds to Cookie

Monster and asked them, “Here is a monster who likes to eat X. What

would you call him?” while varying the X. First he trained them on mass

nouns like mud, which don’t take a plural, until the children would say

mud-eater. That introduced them to the compound construction without

biasing their subsequent answers. Then he asked the children what they

would call a monster who likes to eat rats. The children virtually always

said rat-eater, not rats-eater, even though they had just heard the

experimenter say rats. In contrast, they often called a monster who likes to

eat mice a mice-eater—and those children who occasionally said mouses never used it in compounds, as in mouses-eater. The avoidance of regular

plurals was not simply an aversion to the sound -s inside a compound. As

with adults, when the children were asked about pluralia tantum nouns such

as pants and clothes, which sound regular but have to be stored like

irregulars, they were happy to call the monster a pants-eater or a clothes-

eater. 44

Gordon then tested whether children could have learned the distinction

by noticing irregular plural-containing compounds such as teethmarks in

their parents’ speech, while noticing the absence of regular-plural-

containing compounds such as clawsmarks. He examined all the

compounds in standard frequency counts and discovered that neither kind

of plural-containing compound is common; virtually all commonly used

compounds take a singular first noun, such as toothbrush and mousetrap. Most of the children walked into the lab never having heard a compound

containing a plural, but the first time they faced the temptation they used

irregular plurals and avoided using regular plurals. Children’s sensitivity to

the distinction between mice-infested and rats-in-fested, Gordon concluded,

is a product of the innately specified architecture of their language system,

not a product of tabulating forms in parental speech. Kim and I asked the same question of children’s ability to distinguish

flied meaning “covered with flies” from flew meaning “soared.” Children

hear plenty of verbs-from-nouns, such as to fish, to plug, to rain, and to

screw in. We discovered, however, that they do not hear any verbs-from-

nouns that sound like an irregular verb, such as flied out or high-sticked. 45

That means that prior experience could not have told them what to do when

a verb’s sound calls for one past-tense form and its structure calls for

another; they tend toward the correct answer on their own. Of course, the speech heard by young children must contain information

that tells them that an inflection is regular to begin with. What is that

information?

It cannot be simply the presence of added material on some

words, because that would not distinguish the regular -ed in pat—patted

from the irregular -en in shake-shaken. Nor can it be the sheer number of

words bearing added material, because we saw that children’s use of a rule

has nothing to do with the proportion of regular verbs in their parents’

speech or in their own vocabularies. How can children recognize a regular inflection when they hear one?

Suppose that children’s language systems are prepared for words and for

rules, and are always on the lookout for examples of each in parents’

speech. Children listen for stretches of sound that fit the canonical pattern

for a word in their language and that are arbitrarily paired with a meaning. Those are grabbed by the word system and stored as roots. Children also

listen for words that might be modified versions of some other word after a

rule has had its way with it. Those are snipped into stems and suffixes and

the suffix is stored separately as rule-ready material. To distinguish words

from rules, then, children must have antennae for signs that a pattern of

vowels and consonants has been added to a word rather than being part of

that word. What might those signs be?

They could be the kinds of information that

linguists themselves use to determine whether an inflection is regular, the

kind of evidence we explored in the preceding two chapters. If children hear

a suffixed version of a verb that falls into a family of similar irregular verbs,

such as blinked and showed (which sound like they should belong to the

drink-shrink-sink family and the blow-grow-throw family), they can infer

that the words have been modified by something strong enough to have

nullified the pull toward the family. If they notice suffixed verbs based on

nouns, such as combed and fished, or on onomatopoeia, such as cracked and

squeaked, they could hear the noun or the environmental sound inside the

word, and assume that the residue must have been added by a rule that is

free to apply to words that aren’t verb roots. If children hear suffixed verbs

with nonbasic sounds such as attached and exercised (which are

polysyllabic), they can guess that these forms are unlikely to be roots linked

in memory to other roots; the extra bit is likely to have been added by a rule

that doesn’t care about sound. We don’t know whether children rely on

these telltale signs, but we do know that the signs are available if children

knew what to listen for. All four kinds of verbs may be found in the

vocabularies of young children before the stage at which they clearly apply

rules in errors such as singed. 46 So children could use these signs of wordhood and rulehood if they had the mental apparatus of words and rules

to interpret what they hear. Once a suffix has been identified as a rule

product using the audible cues, it would be available for productive

combination with new verbs. Children’s speech errors, which make such engaging anecdotes in poetry,

novels, television features, and web sites for parents, may help us untangle

one of the thickest knots in science, nature and nurture. When a child says It

bleeded and it singed, the fingerprints of learning are all over the sentence. Every bit of every word has been learned, including the past tense suffix -

ed. The very existence of the error comes from a process of learning that is

still incomplete: mastery of the irregular forms bled and sang. But learning is impossible without innately organized circuitry to do the

learning, and these errors give us hints of how it works. Children are born

to attend to minor differences in the pronunciation of words, such as walk

and walked. They seek a systematic basis for the difference in the meaning

or form of the sentence, rather than dismissing it as haphazard variation in

speech styles. They dichotomize time into past and nonpast, and correlate

half the time line with the evanescent word ending. They must have a built-

in tendency to block the rule when a competing form is found in memory,

because there is no way they could learn the blocking principle in the

absence of usable feedback from their parents. Their use of the rule (though

perhaps not the moment when they first use it) is partly guided by their

genes. They spontaneously deploy their new rule to a wide range of words

coined by an experimenter or by themselves, and to verbs whose irregular

forms are too faint to retrieve. Children fit the rule into its proper place in

the logic of their grammatical system, keeping regular forms out of certain

word structures and irregular forms out of others. I suspect that in other parts of our psychology the interaction of nature

and nurture has a similar flavor: Every bit of content is learned, but the

system doing the learning works by a logic innately specified. Charles

Darwin captured the interaction when he called human language “an

instinctive tendency to acquire an art.” “It certainly is not a true instinct,” he

noted, “for every language has to be learned. It differs, however, widely

from all ordinary arts, for man has an instinctive tendency to speak, as we

see in the babble of our young children; while no child has an instinctive tendency to brew, bake, or write.”47

8

THE HORRORS OF THE GERMAN

LANGUAGE

Though it is sometimes easy for Americans to forget, English is not the

only language spoken in the world. Humans babble in some six

thousand languages falling into thirty-odd families. For many reasons, those

mother tongues are a motherlode for the understanding of language and

mind. First, no one is biologically disposed to speak a particular language. The

experiments called immigration and conquest, in which children master

languages unknown to their ancestors, settled that question long ago. This

means that if some feature of language is the handiwork of a fundamental

mechanism of the human language faculty, it ought to be visible anywhere

from Lapland to Lesotho, from Peru to Papua New Guinea. Also, to understand language we have to test hypotheses about cause

and effect, but linguists don’t have the luxury of synthesizing a language in

a test tube and seeing how it is spoken, learned, and changed. The

differences among languages already out there make up the only laboratory

apparatus that allows a linguist to vary one factor and see how it affects

another. Finally, no one supposes that language evolved six thousand times. We

find different languages because people move apart and lose touch, or split

into factions that hate each other’s guts. People always tinker with the way

they talk, and as the tinkerings accumulate on different sides of the river,

mountain range, or no-man’s-land, the original language slowly splits in

two. To compare two languages is to behold the histories of two peoples:

their migrations, conquests, innovations, and daily struggles to make

themselves understood. What do all languages have in common, and how do they differ?

1 All

languages have a stock of morphemes (word parts) and a set of conventions

for assembling them into meaningful combinations such as complex words,

phrases, and sentences. When words are assembled, they may accept

suffixes, prefixes, and infixes (insertions), may undergo a change of vowels

or consonants, or may be reduplicated: A part is repeated, as in Malay,

where the plural of orang “person” is orang-orang. Words are modified not

only for tense and number but also for person, case (the role of a noun in

the sentence), aspect (how an event unfolds in time), definiteness (the

distinction between the and a), gender (kind), voice (active or passive), mood (indicative, imperative, subjunctive), polarity (true or false), and a

handful of other distinctions. In all languages there are exceptions to some

rules, that is, irregularity. As in English, irregular forms tend to belong to

words that are used frequently. This does not mean, however, that every inflection in every language

has both regular and irregular words. Even in English we find all the

possibilities. Though the plural and the past tense have mixtures of regular

and irregular words, the progressive is completely regular: Even the most

defiant verbs, he, do, and have, accept the -ing suffix without protest: being, doing, having. In contrast, the English word for the inhabitant of a city or

state is completely irregular, with Londoners in London but Bostonians (not

Bostoners) in Boston, Lousianans in Louisiana but Hoosiers in Indiana, and

no one knowing what to call someone from Massachusetts, the Northwest

Territories, or the United Kingdom. Other languages can be even more

extreme. Turkish verb inflection is a combinatorial dream that cranks out

millions of perfectly predictable forms; Russian noun inflection is a

memorizational nightmare whose declensions are riddled with holes. The

Russian humorist Mikhail Zoshchenko wrote a story about a night

watchman who couldn’t order a set of pokers because, like most Russians,

he didn’t know the genitive plural.2

Some languages, such as Chinese, don’t inflect words; morphology

consists of compounding and a few derivations. Others, such as those in the

Bantu and Amerindian language families, assemble words in layer upon

layer. When words are built in steps, regularity or irregularity can infect

each of the steps separately. For example, French verbs fall into three

families, of which the -er family is considered “regular” because it is the

largest and absorbs most of the new words entering the language. Yet some

-er words, such as aller (to go), have many irregularities; conversely, once

you know that a verb belongs to one of the other two classes, you can apply

a regular pattern of conjugation to many of the verbs within that class

(though not all).

In a language that builds words in stages, therefore, it isn’t

meaningful to say that a word is “regular” or “irregular” full stop; some

parts of a word may be regular, others irregular, depending on whether the

part was specified by a general rule or stored with the word’s root.3

If you have studied a foreign language, you know about irregularity all

too well. Everyone dreads the warnings in the tutorials: “NOTE: NOT ALL

VERBS ENDING IN -RE ARE REGULAR. YOU MUST LEARN

WHICH VERBS ARE ACTUALLY REGULAR -RE VERBS AND

WHICH FOLLOW AN IRREGULAR CONJUGATION PATTERN.” It

also is easy to forget which nouns and verbs in a foreign language are

regular and which are irregular. For example, one snatch of Chinese

dialogue in a Hong Kong movie was rendered into the English subtitle:

“Greetings, large black person. Let us not forget to form a team up together

and go into the country to inflict the pain of our karate feets on some ass of

the giant lizard person.” 4

Though every language textbook discusses “regular” and “irregular”

forms, the concept of regularity they have in mind doesn’t always match the

one we have been examining here. “Regular” almost always is equated with

the pattern followed by a majority of words in the language, or with the

pattern adopted by newly coined words. But I have been using it in a

different sense, one that pertains to the mental processes of speakers rather

than to numbers and citations in a dictionary. “Regular” here refers to a rule

that speakers treat as the default: an inflectional pattern they can apply to

any word in a category, even if the word has never been stored with that

pattern, or with any pattern, in memory. According to this theory, a regular pattern could, in principle, apply to a

minority of words in a language, with the majority having to be learned one

by one. (As James Thurber said, “There is no safety in numbers...

or in

anything else.”) And the rule could fail to apply to a new word if the word

is so similar to irregular words in memory that analogy is irresistible (as in

to spling, which most people inflect as splang or splung).

The only way to

know whether an inflection is regular in the psychological sense is to see

whether people apply it when their memory is blocked: when the word is

new, rare, unusual, foreign, rootless, or headless (the circumstances

encountered in chapters 5 and 6).

The difference between the two senses of regular—“majority of words” versus “applied as the default”—uncovers

intriguing aspects of the psychology of language and the history of

language, and shows how one can affect the other. Connectionist pattern associators love frequent patterns, and they have

thrived from the fact that the regular past-tense suffix -ed is “regular” in the

textbooks’ sense of applying to the majority of verbs in English. Pattern

associator networks generalize well when a pattern is spread over many

connections by a large and diverse set of items in the training set. It is no

coincidence that people generalize -ed and that most verbs in English take -

ed, the connectionists say; people, like pattern associators, go with the

numbers. 5

The argument has problems, as we saw in the last two chapters. Children presumably learn from hearing a word used, not from the mere

existence of a word, and irregular verbs such as did, made, and took are

heard more often than regular verbs; the high frequency with which each

one is used makes up for the fact that there are fewer of them. Children fail

to use the plural rule while their noun vocabulary is almost completely

regular (since there are only a handful of irregular nouns in English), and

they use the past-tense rule exuberantly well after the vocabulary spurt in

which the mixture of regular verbs increases most sharply. And the

vocabulary numbers do nothing to help the pattern associators problems

with the regularization of homophones, headless words, and rootless words,

because the models don’t even register these distinctions and so must be

color-blind to them. Still, as long as the regular suffix is both the most common form and the

most generalizable form, we can never tease them apart and definitively

rule out the connectionists’ suggestion that people are driven by the

statistics. We need a language that breaks the confound by having a regular

pattern found in a minority of words. If speakers still applied the pattern to

words that lack associations in memory—that is, if they use it as the default

with rare, novel, unusual, rootless, and headless words—then we would be

sure that it is the kind of mental operation implementing the pattern, not the

prevalence of the pattern, that makes the regular pattern special. Note that this wish—for a language whose regular pattern is in the

minority—is an oxymoron according to the textbook definition of

“regularity” as the majority pattern. It is unexceptional, though, according

to the psychological definition as the product of a mental rule; the

psychological definition says nothing about numbers. The question is: Does

such a language exist?

Could there be a language so perverse, so twisted, so

sadistic, that it inflicts irregular forms on its speakers a majority of the

time?

I quote from Mark Twain’s essay, Die Schrecken der Deutschen Sprache

(the horrors of the German language):

A person who has not studied German can form no idea of what a

perplexing language it is. Surely there is not another language that is

so slipshod and systemless, and so slippery and elusive to the grasp. One is washed about in it. hither and thither, in the most helpless

way; and when at last he thinks he has captured a rule which offers

firm ground to take a rest on amid the general rage and turmoil of

the ten parts of speech, he turns over the page and reads, “Let the

pupil make careful note of the following exceptions.” He runs his

eye down and finds that there are more exceptions to the rule than

instances of it. 6

Perfect!

Twain knew, of course, that “the awful German language” (as his title is

usually translated) is no more awful than any other language for the

children who acquire it as a mother tongue. But many foreigners, he noted,

“would rather decline two drinks than one German adjective.”

In standard (High) German, verbs have three forms: an infinitive, a

preterite or simple past, and a participle: kaufen—kaufte—gekauft “to buy—

bought–(has) bought.” (Mercifully, the simple past is seldom used in casual

speech.) A participle has a prefix, usually ge-, the verb stem, and a suffix,

either -t or -en. The verbs themselves come in three flavors. Weak verbs,

such as kaufen, are regular; they are like English regular verbs such as play

—played. Strong verbs are irregular: The stem usually changes

unpredictably, as in gehen—ging—gegangen “to go—went–(has) gone,”

and they take the suffix -en. They are like English strong irregular verbs,

such as sing—sang—sung. Mixed verbs also are irregular: They take the -t

suffix, but their stem changes unpredictably, as in rennen—ran-nte—

gerannt “to run—ran–(has) run.” They are like English weak irregulars, such as sleep-slept. The parallels are no coincidence. English and German evolved from a

common ancestor, Proto-Germanic, spoken about two to three thousand

years ago. The suffixes spelled -ed in English and -t in German are

descendants of the dental suffix in Proto-Germanic (so called because it w

as pronounced with the tongue against the gum ridge behind the teeth).

The

vowel-change patterns, or ablaut, came from an even earlier ancestor, Proto-

Indo-European, whose conjugations had long since decayed into

irregularity. The parallels between English and German are visible in

cognate verbs that have similar irregular forms, such as singen—sang—

gesungen “sing—sang—sung.”

In German, as in English, the irregular verbs have higher frequencies of

use.7 Among the thousand most common German verbs (which embrace

about 96 percent of the verb uses in a large corpus), the irregulars are used

an average of 640 times in every million words, and the regulars are used an

average of 77 times. (The comparable figures for English are 684 and 73.)

As in English, German irregular verbs come in families. For example,

singen—sang—gesungen resembles sinken—sank—gesunken and trinken—

trank—getrunken; sehen—sah—gesehen resembles lesen—las—gelesen and

geben—gab—gegeben. Membership in the families, however, cannot be

captured by a rule: The singen group has exceptions like beginnen—begann

—begonnen, and the sehen group has exceptions like gehen—ging—

gegangen. Furthermore, German speakers sometimes have muzzy

judgments about their verbs: The preterite of backen can be either buk or

backte, the participle either gebacken or gebackt. 8 All this suggests that in German, as in English, irregular forms are stored in an associative memory,

which can encourage generalization of the patterns to similar new forms. Yet the verbs of German and English do differ in one way: In German

the irregular verbs are more plentiful. Of the thousand commonest verbs in

English, a majority, 86 percent, are regular, but of the thousand commonest

verbs in German, a minority, only 45 percent, are regular. Many

connectionists have pointed to German as a troublesome case for any theory

invoking rules, because it’s not clear whether there is a regular class, at least

by the traditional definition of “regularity” as the majority pattern. 9

But by the psychological definition of “regularity” as the default, the

German weak suffix passes with flying colors. The linguists Richard Wiese,

Harald Clahsen, and Dieter Wunderlich have shown that German -t works

just like English - ed: It goes on any verb, as long as the verb does not have

an association with an irregular root in memory:10

• English speakers apply -ed to rare verbs, such as ablate—ablated;

German speakers apply -t to rare verbs, such as löten-gelötet “welded.”

• English speakers apply -ed to novel verbs, such as wug—wugged;

German speakers apply -t to novel verbs, such as faben—gefabt. • English speakers apply -ed to unusual-sounding verbs such as to

ploamph and to krilg; German speakers apply -t to unusual-sounding verbs

such as qu-ossen and rilken. • In English, -ed can apply to words with irregular homophones, such

as lie—lied and lie—lay or hang—hanged and hang—hung. In German, -t can apply to words with irregular homophones, such as malen—gemalt “to

paint” and mahlen—gemahlen “to grind,” or schaffen—geschafft “to work”

and schaffen—geschaffen “to create.”

• In English, -ed can interlope into irregular-sounding territory, as in

winked and blinked. In German, -t does the same. Weak fehlen—gefehlt “to miss” rhymes with strong stehlen-gestohlen “to steal”; weak kaufen—

gekauft “buy” rhymes with strong saufen—gesoffen “drink booze.”

• English speakers use -ed for onomatopoeia, as in ping—pinged, ding

—dinged, and peep—peeped; German speakers use -t for onomatopoeia, as

in brummen—gebrummt “growl,” flüstern—geflüstert “whisper,” and

klatschen—geklatscht “clap.”

• As with English out-Sally-Rided and high-sticked, German uses -t for

verbs that are derived from other categories and thus cannot have special

past-tense roots listed in memory. These include verbs derived from nouns,

such as frühstücken—gefrühstückt from Frühstück “breakfast,” baggern-

gebag-gert “to dredge” from Bagger “excavator,” and hausen-gehaust “to

house” from Haus “house.” The same is true for verbs derived from

adjectives, such as kürzen—gekürzt “shorten” from kurz “short” and

säubern—gesäubert “to clean” from sauber “clean.” It happens not only

with existing verbs but with new ones made up on the spot: If someone

were to coin a verb gorbatschowen “to Gorbachev,” everyone would give it

the participle gegorbatschowt. • In both languages the weak suffix applies not only to verbs-from-

nouns (which are rootless) but to verbs-from-nouns-from-verbs such as flied

out, which have irregular roots but not in the head position from which an

irregular form can percolate up. For example, the irregular verb halten—

hielt—gehalten “to hold” can be converted into the noun Halt “a hold”

which can be used in the compound Haushalt “household.” The compound

can be turned back into a verb “to housekeep,” but the irregular forms are

unavailable and the regular suffix applies: haushalten—gehaushaltet. • German speakers, like English speakers, are prone to overapplying

the weak suffix (-t) to irregular verbs, resulting in errors such as gesingt for

gesungen. Adults make these errors occasionally; children make them more

often, at rates comparable to English-speaking children. Despite the

prevalence of vowel changes and -en among German participles, German-

speak-ing children seldom overapply them; virtually all their errors consist

of overapplying -t. 11

To verify that flesh-and-blood German speakers, and not just

dictionaries and linguists, apply -t to headless and rootless verbs, Marcus,

Clahsen, Wiese, Ursula Brinkmann, and I ran an experiment in Germany

parallel to the ones we have run in the United States.12 German speakers

were given a questionnaire that asked them to rate participle forms of novel

verbs. Half the new verbs were based on nouns:

Die kleinen dreieckigen Pfeifen für Yuppies sind bei der Kundschaft

gut angekommen. Täglich muβ Tabakhandler Meier die Regale

auffüllen, auf denen die Pfeifen ausgestellt werden. Morgens ist

daher immer seine erste Sorge: \[“The little triangular pipes for

yuppies are a success with the customers. Every day the tobacconist,

Meier, has to fill the cabinets in which the pipes are exhibited. Therefore, his first concern every morning is:”]

Sind die Regale auch schon bepfiffen?

\[“Have the cabinets already been pippen?”]

Sind die Regale auch schon bepfeift?

\[“Have the cabinets already been piped?”]

The other half were existing irregular verbs with a stretched meaning:

Die schöne Ilse glaubt, mit ihrem Pfeifen Karriere beim Film

machen zu können. Wenn sie beim Vorstellungsgespräch gefragt

wird, was sie kann, fängt sie keineswegs an, aus Goethes Faust zu

zitieren. Nein, nein, Ilse beginnt zu pfeifen. \[“Pretty Ilse thinks

she’ll have a career in the movies by her whistling. When asked at

the audition what she can do, she doesn’t start reciting Goethe’s

Faust at all. No, Ilse starts to whistle.”]

Mittlerweile hat sie schon sieben fassungslose Regisseure bepfiffen. \[“Meanwhile, she has already bewhustle seven speechless

directors.”]

Mittlerweile hat sie schon sieben fassungslose Regisseure bepfeift. \[“Meanwhile, she has already bewhistled seven speechless

directors.”]

These items provided a baseline for the popularity of the irregular forms,

and since their meaning has been stretched they bias the experiment against

us and toward the semantic stretch theory (according to which any

extension of meaning repels a speaker from an irregular form).

As predicted, subjects liked the regular forms better than the irregular

forms when the verb was based on a noun, but preferred the irregular forms

when the verb was merely stretched in meaning. Thus in German, as in

English, irregular inflection is confined to words that can be linked to roots

in the mental lexicon. Regular inflection applies to a “verb,” period, and

doesn’t care about the verb’s memory status: It rushes in for words without

roots and for words whose roots are trapped in memory by the absence of a

percolation pipeline in the word’s grammatical structure. So by nine of the tests that establish that -ed is the default past tense in

English, -t must be counted as the default participle in German, despite the

fact that -t applies to a minority of verbs in German. The omnipotence of a

regular rule does not seem to depend on a person’s previously having been

swamped with regular forms. “Does not seem to depend,” I say, because the case is not airtight. It

depends on counting words in the two languages, and counting words is

always a tricky business. Often it is unclear which things should count as a

word, or how many times. First, why stop at the commonest thousand words?

Why not count all of

them?

The reason is that a census of every word ever spoken would turn up

vast numbers of words that only a few people know; certainly no one would

know all of them. An obscure bit of jargon in an anatomy journal can have

no effect on the reader of a fishing magazine and vice-versa. But if you do

go further down the German list, the regular words begin to gain on the

irregulars, because there are only so many irregulars and soon they become

scarce and eventually run out. If you try to inventory every last German

verb—by scanning a database pooled from unabridged dictionaries and

millions of words of text—the regular verbs then become a majority, 78

percent. That’s still lower than the proportion in a large database of English

(95 percent), but the difference between the languages is no longer so clear-

cut. The other perennial problem in counting words is how to tally those that

differ in meaning but share a root. German has many families of verbs such

as ankommen “arrive,” aufkommen “blow up” (the onset of wind or a

storm), and bekommen “receive.” We counted them separately, just as we

counted English stand, withstand, and understand separately, and part,

depart, and impart separately. But that is debatable, because in German the

prefix sometimes can be chopped off the verb and placed elsewhere in the

sentence. As Twain explains:

The German grammar is blistered all over with separable verbs; and

the wider the two portions of one of them are spread apart, the better

the author of the crime is pleased with his performance. A favorite

one is reiste ab—which means departed. Here is an example which I

culled from a novel and reduced to English:

The trunks being now ready, he DE- after kissing his mother

and sisters, and once more pressing to his bosom his adored

Gretchen, who, dressed in simple white muslin, with a single

tuberose in the ample folds of her rich brown hair, had tottered

feebly down the stairs, still pale from the terror and excitement

of the past evening, but longing to lay her poor aching head yet

once again upon the breast of him whom she loved more dearly

than life itself, PARTED.13

If you collapse all the verbs sharing a root, ignoring the differences in their

meanings, then the regularity gap narrows even more: 83 percent of the

German roots are regular, compared to 91 percent of the English roots. 14

What we need is a pattern that applies to a minority of words regardless

of how the words are counted. This time the German language, though

awful to the language student, is helpful to the language scientist. German has five plural suffixes: -en, -s, -e, -er and 0 (zero, or no suffix

at all).

Just to keep learners on their toes, three of the suffixes are

sometimes accompanied by a change in the vowel of the noun, the process

called umlaut we met in chapter 3.

Here are the eight varieties:

German textbook authors have made heroic efforts to impose order on

this mess, but as Twain noted, the counterexamples outnumber the

examples. One linguist eked out ten rules but tacked on seventeen lists of

exceptions. There are some probabilities, such as that masculine and neuter

nouns ending in -ex and -el usually take zero, but nothing reliable. 15 Wiese and Wunderlich argue that seven of the eight plural classes are simply

irregular. The reason that nouns with certain sounds get certain plurals is

not that rules have put them there, but rather that irregular forms are stored

in an associative memory, which makes families of similar forms easier to

remember and encourages people to analogize the plural of one noun to a

similar-sounding noun. The eighth plural, -s, is different. It is by far the least common of the

plural forms, no matter how you count them.16 Among the two hundred

most common nouns, only 1 percent, Autos and Hobbys, take -s. Among the

25,000 nouns in the largest database we could find, only 4 percent take -s. If

you aggressively collapse nouns that share a root, and exclude obscure

nouns that hardly anyone uses, you can boost the proportion to perhaps 9

percent. Even that liberal estimate doesn’t come close to a conservative

estimate of the percentage of nouns taking -s in English: 98 percent. No

quibbles about word counting can alter the conclusion that nouns taking -s

are a large majority in English and a small minority in German. The -s plural is special for another reason: It is regular. 17 Decades ago a German linguist called it the Notpluralendung “emergency plural ending,”

which nicely captures the key trait of regularity in the psychological sense:

It serves as the default, acting whenever memory retrieval comes up empty-

handed.18 The -s in High German is not a historical cousin of the -s in English (unlike -t and -ed), but its similarity to English -s is almost spooky:

• English -s applies to unusual-sounding words and to

words borrowed from other languages; so does German -s. Café, borrowed from French, has a stress pattern unlike that of

German roots, and Kiosk, from Turkish by way of French, is

even stranger. They don’t sound like anything in German, but

they are not left without plurals: -s rushes in, giving Cafés and

Kiosks. • German -s also can set up camp in phonological

territories that are tightly associated with other plurals. It

attaches to nouns that rhyme with irregulars: Schecks

“cheques” despite Flecken “spots,” Labels “labels” despite

Kabel “cables,” Relings “railings” despite Ringe “rings.” That

is, -s is the only plural that can appear with any noun,

regardless of its sound. • In English we talk about Julia and her husband, the

Childs (not the Children).

In German they talk about Thomas

and his wife, die Manns (not die Mann or die Manner).

• We import Renault Elfs (not Elves); they export Opel

Kadetts (not Kadetten).

• Eponymous titles work the same. We enjoy the films

about the Caped Crusader, the three Batmans (not Batmen);

they enjoy productions of the play about the omniscient

alchemist, Fausts (not Fäuste “fists”).

• Both suffixes step in to pluralize quotations. “While

scanning for sexist writing, I found three ’man’s on page 1”

(not men).

“Nach Kor-rekturlesung für sexistische Wortwahl

fand ich drei ‘Mann’s auf Seite 1” (not Männer).

• We use -s with nouns that were converted from other

parts of speech. For example, the linguist George Curme quotes

a complaint about workers loafing on the job: “to obtain

surreptitious smokes and loafs” (not loaves).19 German

speakers also use -s for nouns that were converted from other

parts of speech, as in Wenns und Abers “ifs and buts.” They

even use it with nouns converted from entire verb phrases:

• We would use -s for acronyms and truncations, such as

sysmans and OXes (similar to our use of -ed for truncated

verbs, as in lip-synched).

That is one of its uses in German:

GmbHs “corporations”; Wessis “West Germans,” from

Westdeutsche; Sozis and Nazis “socialists, National Socialists,”

from Sozialist. • English-speaking children say mans; German-speaking

children say Manns. The little Dreikäsehochs overgeneralize -

en more often than -s, but given how few nouns they hear with

-s, it’s remarkable that they overgeneralize it at all. 20

• Perhaps the neatest corroboration of the special status of

-s comes from a circumstance in which it cannot occur: inside

compounds. Recall how English loves to form compounds as

long as they don’t have regular plurals inside them: mice-

infested but not rats-infested, teethmarks but not clawsmarks. German is even more profligate with compounding,

notoriously so. A dubious “German Lesson” circulating on the

Internet gives some examples:

Dog: Barkenpantensniffer

Dog Catcher: Barkenpantensniffersnatcher

Dog Catcher’s Truck: Barkenpantensniffersnatcherwagen

Garage for Truck: Barkenpantensniffersnatcherwagenhaus

Truck

Repairman:

Barkenpantensniffersnatcherwagen-

mechanikerwerker

Mechanic’s

Union:

Barkenpantensniffersnatcherwagen-

mechanikerwerkerfeatherbeddengefixengruppe

Piano: Plinkenplankenplunkenbox

Pianist: Plinkenplankenplunkenboxgepounder

Piano Stool: Plinkenplankenplunkenboxgepounderspinnenseat

Piano

Recital:

Plinkenplankenplunkenboxgepounderof-

fengeshowenspelle

Fathers at the Recital: Plinkenplankenplunkenboxgepounderof-

fengeshowenspellensnoozengruppe

Mothers

at

the

Recital:

Plinkenplankenplunkenboxge-

pounderoffengeshowenspellensnoozengruppenuppenwakers

The -en and -er sounds peppering the compounds are not entirely fanciful,

because in German, as in English, irregular plurals easily appear in

compounds:

Professor en kränzchen “professors’ circle”

Frau en laden “women’s center”

Schwein e stall “pigsty”

Gäns e braten “roast goose”

Büch er regal “bookshelf”

Sozialist en treffen “socialists’ meeting”

This free and easy compounding does not extend to -s plurals, however;

Sozistreffen “socialists’ meeting” and Autosberg “cars heap” sound as

awkward as our rats-infested and clawsmarks, and children refuse to call a

monster that eats cars an Autosfresser. 21

The rulefulness of -s is not perfect. Headless “has a” compounds

(bahuvrihis), which ought to be regularized, stay irregular in German:

Gröβmaul—Gröβmäuler “braggarts,” literally “bigmouths”; Geizhals—

Geizhälse “misers,” literally “thrift necks.” And even among the

constructions that do regularize, we find sporadic counterexamples. There

are plausible explanations, but to allay all fears about the reality of the

Notpluralendung we wanted to show it in action in the minds of live

German speakers. 22

We asked German speakers to rate the eight possible plural forms for

several kinds of made-up nouns. Half of them rhymed with existing

German irregular nouns and should be susceptible to analogies. For

example, Pund rhymes with Hund—Hunde, Pfund—Pfunde, and Grund—

Gründe, so people might well be tempted to pluralize it as Punde or Pünde. The other half did not sound like anything in German, such as Fnöhk,

Pröng, and Plaupf. (Plaupf, by the way, is German for “ploamph.”)

These sounds then were turned into three kinds of nouns whose

structures were made clear to the subjects by the sentences in which we

presented them. A third of the sounds were presented as roots, that is,

ordinary German words. For example:

Ich habe einen grünen KACH gegen meine Erkältung genommen. \[“I have taken a green KACH for my cold.”]

Aber die weiβen KACH sind oft billiger und helfen auch besser. Aber die weiβen KÄCH sind oft billiger und helfen auch besser. Aber die weiβen KACHE sind oft billiger und helfen auch besser. Aber die weiβen KÄCHE sind oft billiger und helfen auch besser. Aber die weiβen KACHEN sind oft billiger und helfen auch besser. Aber die weiβen KACHER sind oft billiger und helfen auch besser. Aber die weiβen KÄCHER sind oft billiger und helfen auch besser. Aber die weiβen KACHS sind oft billiger und helfen auch besser. \[“But the white kachs are often cheaper and work better.”]

We predicted that these nouns would be eligible for regular or irregular

inflection, and the choice should depend on similarity in sound: Nouns that

rhyme with existing irregular nouns should tend to get their irregular

plurals. Another third of the sounds were presented as names, which should

elicit the regular plural, -s. Mein Freund Hans KACH und seine Frau Helga KACH sind ein

biβchen komisch. \[“My friend Hans Kach and his wife Helga Kach are a bit strange.”]

Die KACH versuchen immer, ihre Schuhe anzuziehen, bevor sie

die Socken anhaben. Die KACH versuchen immer, ihre Schuhe anzuziehen, bevor sie

die Socken anhaben. \[etc.]

\[“The Kachs always try to put on their shoes before they put on

their socks.”]

The remaining third of the sounds were presented as if they had been

borrowed from a foreign language:

Die französische “KACH” sieht schwarz am besten aus. \[“The French ‘kach’ looks best in black.”]

Aber eigentlich sehen KACH in jeder Farbe gut aus. \[“But actually kachs look good in any color.”]

These nouns also should get -s, though the tendency should be diluted when

they rhyme with an existing German noun, because that makes them easier

to assimilate to the native stock (just as native-sounding imports to English

such as quit and cost sometimes become irregular).

What happened?

As predicted, the subjects preferred irregular plurals

for the roots, and the preference shrank when the roots didn’t rhyme with

existing irregular nouns—the classic associationist trend of generalization

by similarity that we saw for English in chapter 5.

With the names,

however, the preference flipped: The -s plural sounded better across the

board, whether the name rhymed with an irregular noun or not. And with

the foreign borrowings, the -s plural also shot up compared with the roots. With items that rhymed with German nouns and thus could easily be

assimilated, subjects slightly preferred an irregular; with nouns that didn’t

rhyme, they slightly preferred -s. So -s really is different from the other seven plural forms. The other

seven are irregular, and can be generalized only by analogy to roots. The -s

is regular, and is used as the default, in all the “emergencies” in which

memory and analogy fail: unusual roots, unassimilable borrowings, names,

acronyms, truncations, phrases, and quotations. These emergencies may

strike you as a motley collection of exotic constructions, but that is exactly

the point. The heterogeneity of the constructions, and people’s ability to

apply the regular suffix to them the first time they are faced with the choice,

show that people do not have to be trained to associate the suffix with each

construction separately—the constructions need only be given the mental

label “noun.” And the power of -s to serve as a default even though it is rare

among words shows that regularity cannot depend on a pattern being

stamped into a person’s mind through exposure to a large number of regular

words. Regularity comes instead from the mind’s ability to acquire

symbolic rules: operations that apply fully to any instance of a category. The idea that a class of words can be both regular and in the minority

upends the traditional concept of regularity found in every language

textbook. How could the commonsense notion have been so wrong?

The

reason is that the notion is based on a correlation: The most generalizable

inflection in English is also in the majority. The traditional view assumes a

causal relation: Frequent experience leads to a greater tendency to

generalize. But as with many correlations, the causal arrow can be flipped

around. The “English language” that provides the input to a child did not

float down from the sky. A language is the product of generations of

learners and could reflect, rather than shape, their tastes and propensities. English words may be mostly regular because they are the products of a

rule, not the other way around. In Proto-Germanic, the ancestor of English and German, a majority of

verbs were strong; they were the forerunners of today’s irregular verbs. There was also a precursor of the weak -ed/-t suffix, perhaps a reduced form

of the verb do. Owing to its origin as a freestanding word shuffled around

by the rules of syntax, the suffix kept its habit of promiscuously abutting

against other words regardless of their sound or composition. It became the

suffix of choice for new words that could not easily be associated with the

existing strong classes: words borrowed from other languages, and words

derived from other categories. 23

As it happens, the major growth areas in English over subsequent

centuries were precisely those kinds of words. In 1066 William the

Conquerer invaded England. As noted in the satirical history 1066 and All

That, “The Norman Conquest was a Good Thing, as from this time onwards

England stopped being conquered and thus was able to become top

nation.” 24 The conquest had another important consequence. Norman

French became the language of the government and aristocracy for the

ensuing 150 years, and during that interlude the English language was

flooded with French words, including a large number of verbs. In the

following centuries, English also absorbed numerous verbs from Latin due

to the influences of the Church and of Renaissance scholars, who needed

many words for abstract concepts. By sampling from an electronic

dictionary I have estimated that about 60 percent of English verb roots

came from French or Latin. English also is notorious for verbing nouns:

Another 20 percent of our verbs are converts from the noun category. Both

kinds of verbs, once introduced, had to be regular for grammatical reasons:

They are rootless or headless. A reliance on these lazy ways of forming new words is not all that

surprising. In a “Calvin and Hobbes” comic strip Calvin is taking an exam

that asks him to “explain Newton’s First Law of Motion in your own

words.” He writes, “Yakka foob mog. Grug pubbawup zink wattoom

gazork. Chumble spuzz.” Most of us, though, are not so quick at thinking

up new roots. When a new concept needs a label, we borrow it from another

language we know, or we derive it from some other word, or we use

onomatopoeia, or truncation, or an acronym—all of which, as it happens,

breed rootless or headless words that call for regular inflection. Brand-new

roots coming out of the blue like Calvin’s could take irregular inflection if

they sounded similar enough to old irregulars; for example, Calvin’s z ink

could very well be inflected as zank and has zunk. But such roots, dreamed

up out of nothing, are rare; all but a handful of English roots can be traced

back centuries or millennia. 25

So the traditional definition of regularity, and the connectionist

explanation based on it, have it backwards. It’s not the case that a majority

of English verbs are regular, which trains English speakers to use the

regular suffix as the default. Rather, English speakers and their linguistic

ancestors have used the regular suffix as the default for millennia, and that

is why the majority of today’s English verbs became regular. Nothing

changed in the minds of speakers as regular verbs grew from a minority to

the majority. German also borrowed from Latin and French and derived many verbs

from nouns, though not as much as English did. The German-speaking

lands share a long border with France but they never endured a centuries-

long domination by a French-speaking elite, as Britain did after 1066.

German also did not have to resort to verbing nouns as much as English did,

because it has another way of adding to its verb stock: the prefixed verbs

that Twain complained about. Thus today’s German speakers inherited a

smaller proportion of regular verbs. The difference in statistics is a sheer

accident of history, a consequence of who won the Battle of Hastings. The

psychology of English speakers and German speakers is the same. The psychology of -s is also the same, though its history is very

different. Old English was even more awful than German, with at least nine

ways of making a plural, including the suffix -as, an ancestor of today’s

regular plural. When the vowels at the ends of words eroded to a bland

schwa in Middle English, all the plural forms but -es and -en shriveled. The

suffix with -s prevailed because it was audible, could be pronounced after

both vowels and consonants, and crucially, was imported in great numbers

on plural nouns borrowed from Norman French, which also happened to

use a plural suffix that contained -s. Middle English speakers mentally

merged the French and English -s, and hearing it on all those foreign-

sounding words, reanalyzed it as an all-powerful regular suffix that could

apply to nonroots in general. 26

High German (the standard dialect spoken in the southern and central

parts of the country) has different statistics, because the elevation of -s to

the status of a default rule came much later in its history. 27 In Old and Middle High German, -s was completely absent. It first appeared when

plural nouns bearing -s were borrowed from Low German (spoken in the

north and east), Dutch, English, and French, especially in the eighteenth

century. By the nineteenth century, speakers started to generalize -s to all

borrowed words and other “indeclinabilia.” This offended the prescriptive

grammarians of the era, who called the suffix “strange” and “ignorant” and

urged people to stay away from it. People ignored them, and today -s is

going strong as the regular default. The difference between the number of

regular plurals in English and German is simply a difference in how long

the -ss have been around in the two languages snatching up the new

headless and rootless nouns. As with -ed and -t, their psychology is the

same. With this combination of psychological and historical insight we now can

understand why every language—indeed every inflection in each language

—has a different mixture of regular and irregular forms. Each mixture

arises when unique historical events—conquest, immigration, trade, fads in

speaking—are handled by an unchanging mental tool kit, which contains a

frequency-and similarity-loving associative memory and a promiscuous

combinatorial grammar. A common misconception is that because Old English had more

irregular verbs than Modern English, languages always evolve from

irregular to regular. Languages don’t consistently evolve in either direction,

because different psycholinguistic processes constantly create and destroy

the two kinds of words or convert one into the other. These processes have

made scattered appearances throughout the book; let me collect them into

two lists. New irregular forms arise when:

• A newly coined root is similar in sound to a family of

irregulars and is analogized to them, as in spling—splang—

splung. • An existing regular verb is similar in sound to a family of

irregulars and is analogized to them, as in kneel—knelt and

sneak—snuck. • A rule is rendered opaque or obsolete by changes in

pronunciation habits, and its former outputs are thereafter

memorized as irregulars, as in foot—feet (from a rule-

governed shift in the pronunciation of oo that had originally

been triggered by a plural suffix, since deceased).

• A regular form is slurred in speech, obscuring its anatomy,

as in made and had, formerly maked and haved. • Two words merge and play musical chairs, and the inflected

form of one word becomes a suppletive version of the other,

as in go-went, the outcome of a shakeout after the merger of

go and wend. • A complex word is assembled with an irregular root as its

head, as in became, overate, chessmen, oilmice. New regular forms arise when:

• A newly coined root is unlike any existing word, as in

snarfed and moshed. • An irregular form slips in frequency until people can no

longer recall it on demand, as in chide—chid and geld—gelt. • A word enters the language without a root, via

onomatopoeia ( pinged), eponymy ( the Childs), or borrowing

from another language ( talismans, succumbed).

• A word is converted from a different part of speech and

lacks the right kind of root for the inflection, as in high-

sticked and braked. • A complex word is assembled without a head and hence

without a percolation pipeline, such as bahuvrihi compounds

( lowlifes, sabertooths) and double conversions ( flied-out,

grandstanded).

• A complex word is assembled with a regular root as its head,

as in outwalked and carseats. The lists show how many kinds of brain work go into sculpting just one

corner of a language. German and English are sister languages, so it may not be surprising that

their inflections work the same way. For all we know, the subtle

grammatical phenomena we just explored may be quirks invented by one

Germanic tribe rather than a design feature of the human language faculty. Do other languages show the signatures of rules as well?

Let’s explore a

succession of tongues in an expanding circle from English. 28

Closest to home is Dutch. Like German, Dutch is a West Germanic

language that began to diverge from English in the fifth century A. D. when

the Angles, Saxons, and Jutes left northern Germany for Britain. Like

German and English, Dutch has strong irregular verbs and weak regular

verbs. The irregulars fall into families; for example, most verbs with ij

change the vowel to e and take the suffix -en. 29 But the irregulars are memory-bound, and just as we saw in English, the regular suffix applies

when memory is sealed off by a word’s structure: Verbs made from nouns,

even when they sound like they belong to an irregular class, are claimed by

the regular ending -den. The linguist Chris Collins made up some new Dutch verbs, such as

pijlen “to draw arrows” (from de pijl “arrow”), and vijven “to throw five in

dice” (from de vijf “ five”).

Dutch speakers said their past-tense forms had

to be regular pijlden and vijfden, not irregular pelen and veven. 30

Remarkably, Dutch has two plurals that pass our stringent tests for

regularity, -s and -en. They divide up the territory of noun roots by sound, -s

getting the roots that end in unstressed vowels or vowel-like consonants (Z,

n, and r), -en getting the others. Within their fiefdoms each applies as the

default. A few nouns are irregular, because they either take the suffix - eren,

undergo a vowel change, or take -s when their sound demands -en or vice-

versa. But all revert to the proper regular suffix when they are twisted into

names or quotations. For example, irregular rund—runderen “cows,” kok—

koks “cooks,” and engel—engelen “angels” become regular Rund—Runden

“the Runds,” Kok—Kokken “the Koks,” and Engel—Engels “the Engels.” 31

Now we step out of the Germanic family. French is a descendant of the

“Vulgar” or common Latin of the Roman Empire, and belongs to the Italic

branch of the Indo-European family of languages. Italic and Germanic split

apart some time between 2000 and 1000 B. C. The regular French plural

suffix is -s, but it is now silent in most cases; in spoken French plurality is

generally conveyed by the article, not the noun. But some nouns do have an

audible plural: Nouns ending in -al or -ail take the irregular ending -aux (pronounced ō), such as journal—journaux “newspapers,” hôpital—

hôpitaux “hospitals,” cheval—chevaux “horses,” and travail—travaux

“works.”

Cyrus Shaoul, an MIT student, asked native French speakers to rate

regular and irregular plural forms for a variety of nouns ending in -al and -

ail. The Francophones liked -aux for the familiar irregular nouns that

require it, of course, and they also liked it for unfamiliar nouns that sounded

like other irregular nouns, such as the obscure sénéchal “butler” and the

nonsense grevai. But they flipped their preference to the regular -als in an

impressive number of default circumstances:32

Unusual-sounding

sluzjal—sluzjals (unidentified novel

nouns

objects)

Names

Segal—Segals (actors like Steven

Segal)

Onomatopoeia

spral—sprals

(sounds

of

an

elephant sitting on a motorcycle)

Quotations

“hôpital”–“hôpitals” (instances of

the printed word “hospital”)

Surnames

Cheval—Chevals (Mister Cheval

and his daughter Brigitte)

Product names

Capital—Capitals

(Kawasaki

“Capital” motorcycles)

Eponyms

Arsenal—Arsenals (rap artists like

DJ Arsenal)

Acronyms

Original—Originals

(L’Ordre

Revolutionaire

Iconoclaste

Gaulois pour l’INdependance des

ALsaciens

“the

Gallic

Revolutionary Iconoclastic Order

for the Independence of the

Alsatians”)

Plus ça change, plus c’est la même chose. French falls within the Indo-European family, and a skeptic still might

wonder whether systematic regularization was an invention of the ancient

tribe that begot the family. Hungarian is one of the few European languages

outside the family. That fact, combined with the disproportionate number of

brilliant Hungarian mathematicians and scientists, led one physicist to

suggest that Hungarians are an advanced race of space aliens, but that

theory is no longer believed.33 Hungarian belongs to the Uralic family,

which also includes Finnish, Estonian, Lappish (now called Saami), and the

Samoyedic languages spoken over a vast range of Arctic Russia. The family

descended from a language spoken in the north Ural Mountains more than

7,000 years ago. Hungarian itself is a souvenir of Magyar hordes from the

Eurasian steppe who invaded central Europe in the ninth century A. D. The linguist Edith Moravcsik has made an interesting observation about

irregularity in the language.34 Several Hungarian nouns have a distinctive

set of suffixes, often accompanied by a shortening of the vowel:

When they are turned into names, however, the declension changes:

The new suffixes and the unchanged vowel match one of the regular

declensions in Hungarian, which speakers also apply to borrowed words, as

in tele-fonok—telefonja—telefont. We don’t know how pervasive these

effects are, but it is remarkable to see it at all in a language so distant from

those we have examined thus far. Most linguists think that any traces of a common ancestry between huge

families of languages such as Indo-European and Uralic are lost in the mists

of time. But a few have proposed that Indo-European, Uralic, and Altaic

(which includes Turkish, Mongolian, and Azerbaijani) belong to a

superfamily called Eurasiatic, the legacy of a hypothetical group that

peopled Eurasia toward the end of the last Ice Age 10,000 years ago. What lies outside of Eurasiatic?

The Afro-Asiatic family, formerly

called Hamito-Semitic, originated from a language spoken in the seventh

millennium B. C. and today dominates north Africa and the Middle East. Its

two most famous languages, Arabic and Hebrew, offer not only far-flung

corroborations of the regularization effect but also new evidence that a

regular default is not a by-product of vocabulary statistics. Arabic came from the language spoken by nomadic tribes in northwest

and central Arabia in the centuries around the time of Jesus. The most

common plural in Arabic, the “broken plural,” imposes semi-systematic

changes on the singular: kitābun—kutubun “books,” madrasatun—madārisu

“schools.” Broken plurals apply to families of similar nouns with canonical

patterns of consonants and vowels. The “sound plural,” in contrast, is a pair

of suffixes (masculine -uun and feminine - aat) that apply cleanly to a small,

motley collection of nouns that don’t come from standard roots. The

collection includes proper names, nouns derived from verbs, diminutives

(small or cute versions of things, as in doggie or duckling), unassimilated

borrowings from other languages, and the names of the letters of the

alphabet, which are mostly non-canonical in sound. Examples include

Othman—Othmanuun (a man’s name), Ramadaan—Ramadaanaat (the

month of Ramadan), and tilifuun—tilifuunaat “telephones.” The sound

plural fits the criteria for a regular rule, and as in German, it applies as the

default even though only a few examples trickle into learners’ ears. 35 Just

as remarkably, Arabic-speaking children often overgeneralize the sound

plural, despite its scarcity. 36

Hebrew dates from the second millennium B. C., and for nearly two

millennia after the Roman destruction of the second Jewish kingdom it was

preserved as the language of Jewish scripture and ritual. At the turn of the

twentieth century Hebrew was revived as a living language by Jewish

settlers in Palestine who wanted to shake off all trappings of the ghetto and

shtetl, including the Yiddish of their parents.37

The psychologist Iris Berent has shown that modern Hebrew nails shut

the final escape hatch for the connectionist theory that the generalizability

of regular patterns comes from the statistics of regular words in a language. Several connectionist modelers have replied to our arguments about

German and Arabic by saying that it may not be the number of regular

words that is critical so much as the scattering of regular words in

phonological space. Suppose irregulars fall into tight clusters of similar

forms ( sing, ring, spring; grow, throw, blow; and so on), while regulars are

kept out of those clusters but are sprinkled lightly and evenly throughout

no-man’s-land ( rhumba’d, out-Gorbachev’d, oinked, and so on).

Then one

can design pattern associators that devote some of their units and

connections to the no-man’s-land, and they will deal properly with any

subsequent strange-sounding word.38 These models cannot be taken

seriously as theories of a human child, because they have the inflections of

a language innately wired in, one output node per inflection, and merely

learn to select from among them. And as usual, the problem of rootless and

headless words is ignored. It is interesting nonetheless to test the general

idea that certain patterns of clustering among regular and irregular sounds

are necessary for people to generalize the regular inflection freely. In Hebrew most masculine nouns are pluralized with -im, such as bul—

bulim “stamps”; most feminine nouns are pluralized with -ot, such as mora

—morot “teachers.” But about two hundred nouns are irregular and take the

plural of the other gender, such as masculine kir—kirot “walls” and

feminine dvora—dvorim “bees.” By now you should be able to predict what

will happen when Israelis are asked to choose the best plural for irregular-

sounding nouns:39

In my friend’s room, the kirot/kirim are covered with paintings. The kir is a French drink. To prepare two kirot/kirim. mix two

glasses of champagne and a quarter glass of Cassis liquor. My French friends Brigitte and Jean Kir arrived for a two-week

visit. The Kirot/Kirim will stay at my house during the first week. Speakers know that irregular kirot is correct for the basic word meaning

“wall,” but switch to regular kirim for foreign words and names. They also

apply regular -im to unusual-sounding made-up nouns, such as tcharlak,

krazastriyan, and gogof; indeed they applied it to the strange nouns as

readily as to nouns that were close in sound to existing regular nouns. Regular plurals in Hebrew are picture-perfect examples of a default rule. Here is the punch line: Regular and irregular nouns live cheek-by-jowl

in the same phonological neighborhoods. Irregular nouns do not carve out

their own distinctive sounds, as in English sing—sang, sink—sank, drink—

drank or German singen—gesungen, sinken—gesunken, trinken—

getrunken. Most irregular nouns have sounds that are stereotypical of

regular nouns. For example, the irregular kir—kirot “walls” squats as a one-

word minority in a neighborhood dominated by thirty-one regular nouns,

such as kis—kisim “pocket,” min—minim “gender,” pil—pilim “elephant,”

and shir—shirim “song.” Similarly, irregular zanav—znavot “tails,” valad—

vladot “newborns,” and three other irregulars pepper a space filled by forty-

three regular nouns, such as barak—brakim “lightning” and marak—

mrakim “soups.” Irregular nouns are so well interspersed with their

neighbors that no one can draw a line putting them on one side and the

neighbors on the other. And that cramps the latest hope of connectionism to

explain away rule-governed generalization as a by-product of the statistics

of the input. Believe it or not, even Hebrew and English may belong to a discernible

family with a common ancestor. A few dauntless linguists believe that

Eurasiatic, Afro-Asiatic, Dravidian (the languages of south India), and

South Caucasian form a superfamily called Nostratic. The Nostratic

speakers would have been a group of hunter-gatherers who originated in the

Middle East and spread through Europe, Northern Africa, and all but the

eastern part of Asia about 15,000 years ago.40 To peer outside this

superfamily, and give the words-and-rules theory one more hurdle, we can

look to Chinese. Chinese is a set of languages (we myopically call them dialects) in the

Sino-Tibetan family, which also includes Tibetan and Burmese. It is famous

for having no inflection whatsoever: A word keeps its sound, no matter how

it is used. Some people interpret this as a refutation of any theory in which

inflection is part of a universal design for language. In a message posted to

an internet discussion group for child language researchers, one critic of the

words-and-rules theory, alluding to the lack of inflection in Chinese, asked

sarcastically, “What the hell do Chinese speakers do with their grammar

morphology genes or their dedicated neural mechanisms for regulars vs. irregulars?” The rhetorical question misdescribes the theory: It’s not regular

and irregular inflection per se that are thought to be biologically

distinguishable, but combination and lookup more generally. Combining

morphemes to form a word is morphology; combining words to form

sentences is syntax. Chinese does not have much morphology (it does have

some, in the form of compounding and certain derivations), but of course it

does have syntax, so it does have words and rules. Better still, some of its

rules work as a default, just like the ones half a world away in the Indo-

European, Uralic, and Afro-Asiatic families. In Mandarin Chinese, you can’t talk about a pen or some dogs; you

have to use a classifier or measure word, as in yi-zhi gangbi “a rod of pen”

or yi-qun gou “a pack of dog.” English speakers sometimes have to do the

same; we say a blade of grass (not a grass), a piece of fruit, a strand of hair, a slice of bread, a stick of wood, a sheet of paper, and thirty head of

cattle. Chinese speakers always have to pick a classifier when they want to

refer to a number or an amount of something. Each classifier in Chinese tends to go with a kind of object. There are

classifiers for people, animals, flat things, long flexible things, small things,

one of a pair, and so on. Yet the associations are imperfect and must be

memorized; they cannot be captured by rules. Tiao often is used for long

flexible objects, such as fish, strips of paper, and pants, but it also is used

for shorts and for news items, which are not long and flexible, and it cannot

be used for a strand of hair, which is long and flexible. When Chinese

speakers haven’t memorized the classifier for an object, they use the

classifier for a similar object. This is all familiar from the irregular verbs

and suggests that people store the nouns that go with each classifier in an

associative memory. The linguist James Myers has pointed out that one classifier is

different.41 Ge is used in a hodgepodge of situations that have nothing in common but an inability of the speaker to draw upon memory or an analogy

with something in memory. Ge is used with objects whose size and shape

don’t fit with any classifier, such as xigua “watermelon.” It is used with people who don’t deserve the respectful tone of the classifier for humans,

such as xiaotou “thief” and pozi “hussy.” It is used with abstractions, such

as xiwang “wish” and guojia “country.” It is used with nouns that have been

converted from verbs, such as zhongliao “completion” and tiyan “learning

from experience.” It is used with quotations, as in the real-life example

“You’re good to me, I’m good to you, this- ge ‘good’ comes to have

vitality.” Combitations of ge- and a noun tend to be lower in relative

frequency than combinations of other classifiers and their nouns. People use

ge when they can’t remember a noun or can’t remember its classifier, and

children overgeneralize it to inappropriate nouns. Ge has all of the powers of a regular inflection, though it appears in a

language that has no inflection. Apparently a rule assigns ge to anything

thought of as “noun” unless the noun already has a classifier. Myers’s

analysis of Chinese shows that the bewildering variation among languages

can be misleading. Beneath the variation lie deep universals rooted in the

nature of mental computation. How far from English can we go and still find the fingerprints of rules?

New Guinea was settled more than 40,000 years ago by an extraordinary

group of early modern Homo sapiens who somehow crossed fifty miles of

open ocean to get there. Over tens of millennia they fanned out into the

isolated valleys of the highlands and splintered into tribes speaking more

than 800 languages unrelated to anything spoken anywhere else on the

planet. Most New Guineans had no contact with the rest of the world until

the 1920s and 1930s, when prospectors, traders, and anthropologists began

to explore the interior. Around that time Margaret Mead and her second husband, the

anthropologist Reo Fortune, studied a tribe called the Arapesh. She focused

on gender in the sense of sex differences; he focused on gender in the sense

of inflectional morphology. Mead’s research has not held up well. She

referred to the people as “the gentle Arapesh,” and it turned out that the

men were headhunters. Fortune’s research has held up better, and the

linguist Mark Aronoff recently revisited it using the tools of modern

linguistic theory. 42

Arapesh has thirteen genders. This is not as kinky as it sounds; to a

linguist gender means “kind,” as in the related words genus, generic, and

genre. In Arapesh most of the genders are phonological. The nouns in a

gender end in a particular syllable or phoneme, such as ag, or r. But one

gender is different. Aronoff calls it the “default gender,” and notes that it is

used whenever the gender of a noun phrase “cannot be determined for

whatever reason.” Sound familiar?

One reason an Arapesh speaker may be unable to determine the gender

of a phrase is that the phrase has no head because the noun is omitted, a bit

like the English This is nice or Which do you want?

A second reason is that

the noun phrase is a conjunction of two nouns with different genders

(somewhat like the girl and the boy).

Conjunctions are headless, and when

two nouns have different genders often it isn t clear (in any language) which

noun should pass its gender up to the whole phrase. A third reason is that a

noun may have an unusual sound pattern. In all of Arapesh only two nouns

end in b. They don’t match any of the other genders, and they get thrown

into the default gender. A fourth reason comes from the two genders that do

have something to do with sex; one has nouns referring to human females,

such as barahoku “granddaughter,” and one has nouns referring to human

males such as araman “man.” Words that designate people in a sex-neutral

fashion, such as arapeñ “friend,” ašukeñ “elder sibling,” and batauiñ

“child,” don’t fit either gender and get tossed into the default gender. A fifth

reason is that the phrase may be headed by a sexless pronoun such as “I” or

“you,” which don’t have a gender to percolate up to the phrase as a whole. Once again we see the modus operandi of a rule. In the first seven chapters we explored the tracks and traces of a rule in

action, but only for two suffixes in a single language. In this chapter we

have spotted them in eight other languages ranging from the closest siblings

of English to the most distant strangers. I don’t mean to suggest that all languages work just like English or that

they all can be explained in a simple way by the words-and-rules theory. Every construction in every language throws up a welter of complications

and counterexamples and deserves a book of its own. It is striking, however,

to sight rules living in the same sets of habitats—rare words, unusual

words, headless constructions, converted words, children’s errors—in so

many historically unrelated languages. To see these deep parallels in the

languages of the French and the Germans, the Arabs and the Israelis, the

East and the West, people living in the Age of the Internet and people living

in the Stone Age, is to catch a glimpse of the psychic unity of humankind. 9

THE BLACK BOX

Engineering students sometimes are given the problem of deducing the

design of a circuit in a box from a list of its inputs and outputs. For

decades that has been a pretty good definition of psychology. Though no

one doubts that our thoughts and feelings are caused by the activity of the

brain, that activity has been hard to study because most people don’t want

to hand over their brains to science until they are dead. We infer what parts

of the brain must be doing by seeing what the whole person does when

presented with inputs such as pictures or words or instructions. In the

previous eight chapters I have argued that the brain has different

subsystems for words and rules, not by peering beneath the skull but by

seeking the best explanation for how people speak, understand, learn, and

react to words and sentences. Now the black box is being opened, not with a scalpel but with new

technologies that allow us to see the living brain without invading it. Neuropsychologists have long studied patients with brain injuries and

documented what they could no longer do. Until the advent of

Computerized Axial Tomography (CT or CAT scans) they had to wait for

the patient to die and be autopsied before they could learn what part of the

brain had been damaged. Still newer techniques such as Positron Emission

Tomography (PET) and functional Magnetic Resonance Imaging (fMRI)

give pictures of the workings of the brain, not just its anatomy. DNA testing

is beginning to pinpoint the genes responsible for inherited psychological

conditions and someday will show how they affect the developing brain. This revolution has led to a new field, cognitive neuroscience, and to the

declaration by President George Bush that the 1990s were to be known as

the Decade of the Brain. If words and rules are the ingredients of language, we should be able to

tell them apart in the brain. Parts of the brain that handle memory for words

should be implicated in the use of irregular forms, and parts that handle

rules should be implicated in the use of regular forms. This gives us another

way to test the theory that rules step in when memory fails. Direct

neurological damage to the memory system joins other kinds of memory

failure, such as rare words, unusual words, headlessness, rootlessness, and

childhood inexperience as a circumstance that should summon a rule. Moreover, because regular and irregular forms are so well matched—they

have the same meaning (pastness), the same grammar (tensed), and the

same complexity (one word long)—any difference in how the brain handles

them can help map out the linguistic brain more finely. In the past few years every major technique in cognitive neuroscience

has been applied to the debate on regular and irregular inflection. This

chapter offers a tour of the techniques and what they show about the neural

seats of words and rules. It would be nice if we could pinpoint a patch of brain devoted to rules and

another patch devoted to words, but that fantasy will never come true. The

brain is simply not the kind of organ in which a function has to be carried

out by a chunk of tissue with a recognizable shape. The kneecap has to have

a certain 3-D shape to be a good kneecap, but a sense of direction or a

faculty of emotional intelligence or a language instinct does not. The brain

is the organ of computation, and a computational system cares about how

information flows within it, not about how the system takes up space. In

computers a program or file may end up in different parts of the memory or

disk when loaded onto two machines or onto one machine on different

occasions, and it may be fragmented across far-flung regions of the disk or

memory. As long as the information is preserved and the regions are

properly linked, the program can work perfectly, even though we can never

draw a circle around the part of the memory or disk that contains it. The brain is not a digital computer, of course, but that only makes the

point more strongly; its circuits are not plugged into slots in a motherboard

but somehow find suitable homes in the cerebral cortex as the brain

develops. A mental process is a set of computations in the millions of

synapses of an intricately structured neural network. A network could be

contorted into all kinds of stripes, polka dots, or squiggles on the surface of

the brain and still do the same thing, as long as the neurons were connected

properly. A second reason to doubt we will ever find a rule center and a word

center is that neither words nor rules just dump their products into the

atmosphere. They are part of a complex system and depend on connections

to each other and to many other brain systems. Anyone who has recently

installed a software package on a personal computer knows that even a

simple program needs hundreds of files scattered all over the machine to

coordinate the program with memory, input, output, and other programs. And language has even more features than the latest bloat from the software

industry. The rule system in the brain must be an octopus with tentacles

extending to the mouth, throat, and diaphragm (for speaking), to the ears

(for understanding), to short-term memory (to hold the beginning of a

sentence in mind while figuring out the end), to concepts of every kind (to

plug meaningful words into the sentence), and to systems for reasoning and

planning (to decide what to say and how to say it).

In a man-made device

each component can be in its own box and connected to the others by

cables, but in the brain the system is likely to be a web of interdigitated

blobs laid out along a wide swath of cortex. Think, too, of what it would take to see one of the blobs. In a fantasy we

might imagine that a person could enter a Zenlike trance in which he quiets

his entire brain and thinks nothing but pure past-tense thoughts, making the

putative rule circuit glow. In reality we have to ask the person to do

something, like answer a question or produce a word in response to a cue. Even a task as simple as converting walk to walked forces a person to

remember the instructions, read or listen to each stem, hold it in short-term

memory, send the “past tense” request to the rule system and the lexicon,

activate the rule, suppress false matches with memory if there are similar

irregulars, get the right suffix, join it to the stem, smooth out the sound of

the junction, prepare the sequence for speech, and move the muscles,

possibly while monitoring for errors at every step. A neurological patient

might be unable to come out with a past-tense form if any of these abilities

has been compromised. With all systems in working order, a healthy brain

in a scanner might light up like a Christmas tree. So our ability to tie the steps of language processing to circuits of the

brain is still rudimentary. For now we must settle for something simpler:

clues that regular and irregular words depend on different sets of brain

systems (as well as some in common), and clues that irregulars depend

more on the system for word memory and regulars more on the system for

rules. The human brain is a vast territory: billions of neurons connected by

trillions of synapses in dozens of lumps, sheets, and strands, all packed into

a convoluted three-dimensional shape. But we can orient ourselves, and aim

at the likely habitats of words and rules, by looking at the three major

canyons that carve up the cerebrum. First, the brain has two hemispheres, and in most right-handed people,

language, particularly grammar, is mostly in the left.1

Second, the central sulcus (fissure) subdivides each hemisphere in two. Its front bank is the motor strip, which controls movement. The motor strip

is often drawn in psychology textbooks with shrunken or blown-up body

parts pasted along its length, showing which patch controls which part of

the body. In front of the motor strip lies the rest of the frontal lobe, which

carries out the prerequisites to action: planning and organizing movements,

making decisions, juggling items in short-term memory, directing attention,

executing chains of reasoning, and following goals under the influence of

the emotions. The rear bank of the central sulcus is the somatosensory strip, which

registers touch. The somatosensory strip also is commonly depicted with

severed body parts, showing which patch of brain monitors which patch of

skin. Behind the somatosensory strip, extending in a sweeping curve around

to the parietal, occipital, and temporal lobes, are areas devoted to the other

major senses, vision and hearing. Here we find not only the first stops in the

cortex for the sensations, but also many areas that organize the sensations

into a coherent perception of events in space and time and match

components of the events to concepts of faces, people, sounds, places, tools,

and living things. The third major cleft is the Sylvian fissure, which divides the temporal

lobe from the rest of the brain. The Sylvian fissure anchors the major

language areas, which hang off both banks. Above the fissure toward the

front lies Broca’s area (actually a set of areas), thought to be involved in the

planning of speech, in verbal short-term memory, and in the comprehension

of complex sentences. Below the fissure toward the rear lies Wernicke’s

area, thought to be involved in connecting the sounds of words with their

meanings. A swath of cortex from the lower part of the parietal lobe

sweeping through much of the temporal lobe seems to hold words and their

meanings, with the meanings of words of different categories (colors,

animals, tools, and so on) concentrated in different parts. The division of

language into a front part for grammar and speech planning and a back part

for words and speech understanding is surely too simple. There are hints,

for example, that some verb meanings have links to the frontal lobes, and

that understanding sentences with subtle syntax may involve the anterior

superior (front top) part of the temporal lobe. But it is a reasonable first

cut. 2

The three clefts provide compass points showing us where we might

look for the neurobiology of regular and irregular inflection. If regular

forms, especially rare and new ones, are processed on the fly by rules, we

might find them computed in the anterior (frontal) portions of the left

Sylvian cortex. If irregular forms are stored as words, we might find them

retrieved from the parietal and temporal portions of the left Sylvian cortex. The outsize human brain is a vulnerable organ which can be damaged by

tumors, infections, malnutrition, blocked or burst arteries, and injuries from

falls, bullets, and car accidents. Many people who have suffered these

tragedies participate in experiments that assess what they can and cannot

do. Some do it for money, some to gain insight into what part of them has

been lost and what remains, some as an altruistic contribution to science. When a patient with a brain injury can no longer do something, it is

tempting to conclude that the damaged part of the brain must be the neural

center responsible for the feat the patient can no longer do. But that

reasoning is unsound. Suppose a patient with a damaged X can no longer

name fruits but can still name vegetables. That does not imply that X is the

brain center for the names of fruit. Perhaps naming fruits, for whatever

reason, is more demanding than naming vegetables, and a brain running at

less than full capacity stumbles on the harder task. Decades ago the

neuropsychologist Hans-Lukas Teuber pointed out that links between brain

and mind ought to be based on a double dissociation, involving two kinds

of patients and two kinds of tasks. In our example we would need to show,

at a minimum, that patients with damage to area X have more trouble

naming fruits than vegetables, and that patients with damage to area Y have

more trouble naming vegetables than fruits.3 This doesn’t prove that X is for fruit and Y is for vegetables, but it does suggest that the two areas differ

in the kind of work they do, not just in the amount of work they do, and that

the difference in kind has something to do with the difference between

fruits and vegetables. One famous double dissociation in language involves regular and

irregular spelling in printed words. Some patients mispronounce irregular

words such as yacht and aisle (rhyming them with matched and basil) but have no trouble with nonwords such as wug and dax, whose pronunciations

can be deduced from regular rules of spelling. Other patients, with damage

to different parts of the brain, have the opposite problem: They can

pronounce yacht and aisle, but have no idea what to do with wug and dax. The natural interpretation is that the brain contains two routes from print to

sound. One uses rules, such as “Pronounce the letter pair ee as the sound

’ē,’” and it is needed for new and rare words, which cannot be retrieved

from memory. The other memorizes entire words and their pronunciations,

such as “The string aisle spells the word aisle, which is pronounced ’īl,’”

and it is needed for irregularly spelled words, which defy the rules. The first

kind of patient, with surface dyslexia, has suffered damage to the whole-

word pathway; the second, with phonological dyslexia, has suffered damage

to the rule pathway. 4

The double dissociation gives neural reality to a distinction we might

have guessed on purely computational grounds, and it challenges

connectionist models, which, as with the past tense, try to capture regular

and irregular forms in a single pattern associator memory. Of course

spelling rules are different from rules of grammar: They are consciously

taught and learned, and they show little of the abstract logic of grammar

explored in chapter 6.

But connectionist theorists treat them the same, so any problems they have with models for reading aloud carry over to their

models for grammar. Advertisements for pattern associators boast that regular and irregular

associations are smeared across a single set of connection weights,

eliminating the need for separate boxes for rules and exceptions. The

problem then is how to deal with a double dissociation, such as that

between patients who can no longer read novel words and patients who can

no longer read words with irregular spellings. Generally modelers simulate

brain damage by removing or weakening connections at random, and that

leads to a single dissociation in which the model can no longer handle

irregular words.5 This happens because each irregular form depends on a

rather small number of strong connections between particular inputs and

particular outputs, making the irregular form vulnerable to damage, whereas

regular forms are computed by a diffuse set of many weaker connections,

offering redundancy and resistance to damage. The double dissociation

suggests that the appeal to the aesthetics of a single mechanism may be

misguided; the brain appears to have more than one part. The connectionists reply that sometimes connections in a pattern

associator spontaneously segregate into bundles that concentrate on regular

or irregular associations. As a result, when a model is deliberately damaged

at random to simulate a brain lesion, it may have more trouble with regular

associations on some simulation runs and more trouble with irregular

associations on other runs. But the modelers John Bullinaria and Nick

Chater have shown that double dissociations occur only in artificially small

toy models, where there simply aren’t that many connections for the regular

associations to be spread over; a small amount of damage can therefore hurt

the regular associations as badly as the irregular ones. In any model with a

more realistic number of connections the regular association is distributed

more evenly across the connections, and a simulated lesion always hurts

irregulars more. Bullinaria and Chater conclude that Teuber’s logic of

double dissociation is still sound. 6 In fact the logic is even sounder when

the dissociation can be predicted beforehand from an understanding of what

different parts of the brain do. That ensures that the double dissociation is

not just a fluke, hand-picked after the fact from a mass of random data

going every which way. Michael Ullman, Greg Hickok, Marie Coppola, and I teamed up with

the neuropsychologist Suzanne Corkin and the neurologists John Growdon

and Walter Koroshetz, who study a variety of neurological patients. 7 We began by seeking a double dissociation in regular and irregular inflection in

different kinds of aphasia. Aphasia is an impairment of language following

an injury to the brain, and much of our knowledge of the organization of the

language areas has come from comparisons among different types of

aphasia. 8

Agrammatism is a symptom of some forms of aphasia in which a

patient has difficulty assembling words into phrases and sentences, putting

the right grammatical suffixes onto their stems, and understanding complex

sentences. It frequently appears after extensive damage to the anterior

(front) regions of the language areas around the Sylvian fissure, including

Broca’s area, Agrammatic aphasics usually have trouble with single words

as well, but the trouble often is less severe than their trouble with phrases

and sentences. An agrammatic patient’s speech often sounds like this: “Son...

university...

smart...

boy...

good...

good,” or “Lower Falls...

Maine...

Paper. Four hundred tons a day!” Agrammatic aphasics often

remember “words” in the sense of listemes: memorized chunks that may be

more than one word long, such as “Fit as a fiddle and ready for love.” 9

Agrammatics have great difficulty with grammatical suffixes, usually

leaving them out altogether (particularly in a language like English, where

bare stems are used in the infinitive and the present tense) or using the

wrong one. When reading a list of words, for example, they might read

smiled as “smile” and wanted as “wanting.” Two previous studies, one by

Oscar Marin, Eleanor Safran, and Myrna Schwartz, another by William

Badecker and Alfonso Caramazza, had shown that patients with impaired

grammatical processing make fewer of these errors when reading irregular

past-tense forms and plurals. Our group replicated the effect with a new

sample of five agrammatic patients. The explanation is that regularly

inflected words ordinarily are parsed by rules as they are read, and

agrammatic patients have damage to the machinery that does the parsing. Irregular verbs are matched against memory as wholes, which the patients

can still do. A person who has suffered brain damage could have trouble with

regular forms for reasons other than their regularity. To ensure that their

agrammatic patients don’t simply have trouble pronouncing an -s or -ed at

the end of a word, Marin and his collaborators compared regular plurals

with pluralia tantum, which have to be memorized as irregulars even though

they bear the plural suffix -s. They compared clues with news, buds with suds, and misers with trousers. To ensure that the patients don’t just stop

reading from left to right as soon as they get to the end of a recognizable

word (which would give them smile from smiled), Badecker and Caramazza

gave their patients uncommon words that contained common words, such as

yearn (which contains year), dogma (which contains dog), and pierce (which contains pier).

To ensure that the patients don’t simply have trouble

with the less common or harder-to-pronounce past-tense forms, our group

matched each irregular form with a regular form that had a similar ending

and the same frequency in the language. For example, we matched slid with

tied, swept with slipped, and bought with stayed. Even with all these controls, these patients had greater difficulty reading the regular forms. Reading an inflected word aloud is different from generating it oneself,

and to test patients’ ability to generate past-tense forms, Ullman made up a

battery of items of the form, “Every day I dig a hole; Yesterday I\_a hole.”

Patients read the items or listened to them, and were asked to fill in the

blanks. The verbs were regular, irregular, or nonsense words like spuff and

plam (in other words, a wug-test).

Portions of the lists of regular and

irregular verbs were matched for frequency and for the sequence of

consonants at the end. Rather than testing a large group of patients given a

diagnostic label such as “Brocas aphasia” (which often lumps together

patients with huge messy lesions and a hodgepodge of symptoms), we did a

case study of a patient whose lesion was confined to anterior regions of the

brain and the basal ganglia. His symptoms of agrammatism were

unmistakable, but his ability to name things, though worse than control

subjects, was reasonably good. That suggests that his mental grammar was

more impaired than his mental dictionary, and as we predicted, he had far

more trouble inflecting regular verbs than irregular verbs, was almost

incapable of inflecting novel words like plam, and never overgeneralized

the rule to irregular verbs, which would have resulted in errors like digged. The other half of the double dissociation comes from patients with

anomia, a difficulty in retrieving and recognizing words despite fluent and

generally grammatical speech. Anomic patients often have their words

stuck on the tip of the tongue, and they resort to circumlocutions, pronouns,

and generic words such as something and stuff. Here is a transcript of one

anomic patient trying to name some objects:

\[A clock:] Of course, I know that. It’s the thing you use, for

counting, for telling the time, you know, one of those, it’s a...

\[But

doesn’t it have a name?] Why, of course it does. I just can’t think of

it. Let me look in my notebook. \[His elbow:] That’s the part of my body where, my hands and

shoulders, no, that’s not it. No, doctor, I just can’t get it, isn’t that

terrible?

\[A wallet:] This is a kind of bag you use to hold something; you

may hold materials in it and keep it in your pocket. 10

Anomia is often associated with extensive damage to posterior parts of

the brain, especially the junction of the parietal and temporal lobes, and

often with damage to large parts of the temporal lobe as well.11 Sometimes

patients with posterior lesions have jargon aphasia in which they speak in

their own neologisms, such as nose cone for phone call or words that no

one recognizes at all. Interestingly, they often stick regular suffixes onto

their jargon, a selfadministered wug-test. One patient, struggling to name a

box of matches, said, “Waitresses. Waitrixies. A backland and another bank. For bandicks er bandicks I think they are, I believe they’re zandicks, I’m

sorry, but they’re called flitters landocks.” He does it with verbs as well as

nouns: “She wikses a zen from me,” “He mivs in a love-beautiful home.” 12

This suggests that regular inflection may be computed in a part of the brain

that is distinct from the parts in which words are handled. We tested six aphasic patients with anomia, but focused on one whose

lesion was confined to the posterior parts of the brain. This picture shows

the approximate size and shape of the lesion of the anomic patient and, for

comparison, of the lesion of the agrammatic patient discussed earlier. Just as one would expect if anomic patients suffered greater damage to

their mental dictionary than to their mental grammar, they had more trouble

inflecting irregular verbs than regular ones, were relatively good at

inflecting novel verbs like plam (as much as 80 percent of the time), and

interestingly, made overgeneralization errors like digged, just as children

do. For example, the patient with the circumscribed lesion made the error

25 percent of the time. In these three symptoms, the anomic patient is the

mirror image of the agrammatic patient. The psychologists William Marslen-Wilson and Lorraine Tyler have

doubly dissociated words from rules in the brain in a different way. Recall

from chapter 5 that when intact people hear a word, they are primed to recognize related words. After hearing swan, for example, people recognize

goose more quickly, presumably because the mental dictionary entries

overlap or are linked. According to both the pattern associator theory and

the words-and-rules theory, a pair of irregular words such as find and found

should be associated in a similar way, and sure enough, in experiments

where the words are spoken, found primes find just as swan primes goose. Regular walked primes its stem too, but according to the words-and-rules

theory it is for a different reason: The brain unconsciously analyzes walked

into walk and -ed, and the stem walk primes itself. We know the priming is

caused by grammatical relatedness, not mere overlap in sound, because

overlapping but unrelated words such as gravy and grave do not prime each

other. If regular and irregular priming work in different ways in the brain,

different neurological patients might show priming by regular forms but not

irregular forms, and vice-versa. The technique does not require the patients

to speak aloud (they just press a button if the item is a word), so it bypasses

any remaining worry that irregular forms are more easily pronounced than

regular ones. Marslen-Wilson and Tyler discovered two agrammatic patients in whom

walked did not prime walk (regular inflection), though found did prime find (irregular inflection), and swan primed goose (semantically related words).

Presumably the patients’ circuitry for grammatical analysis was impaired,

so walked and walk struck them as no more related than gravy and grave. But the associations in their mental dictionaries, such as swan to goose,

were not as impaired, and by the same token neither were the associations

between found and find. In a third patient the dissociation went the other

way: walked primed walk but swan failed to prime goose, and as expected, found failed to prime find. The patterns of damage in the patients’ brains were diffuse and hard to

delineate, so we cannot use the double dissociation to identify the brain

areas normally responsible for regular and irregular priming. The first two

patients, who lost priming of regular verbs, had massive damage to the left

hemisphere but no damage to the right hemisphere. Presumably they lost

the areas in the left hemisphere responsible for grammatical processing, but

retained a partial knowledge of words and their relationships in the right

hemisphere. (In healthy people words presented to the right hemisphere can

often prime other words in the same category, suggesting that words and

their meanings are stored in the right hemisphere in addition to the left. 13)

The third patient, who lost priming of irregular words and semantically

related words, had extensive damage to the right hemisphere and patchy

damage to the left. Perhaps both copies of his mental dictionary were

damaged, but just enough grammatical machinery survived in the left

hemisphere to analyze the regular forms. Whether or not this anatomy is

correct, it is clear that regular and irregular verbs depend on different sets of

areas of the brain. Not all brain damage comes in the form of a lesion from a stroke. Neurode-

generative diseases, the result of genes, aging, viruses, autoimmune attacks,

environmental toxins, and unknown causes, can affect some parts of the

brain more than others. The most common neurodegenerative disorder is

Alzheimer’s disease, which strikes about a tenth of people over sixty-five

and half of those over eighty-five. In Alzheimer’s disease, deposits called

plaques accumulate around neurons, and tangled filaments accumulate

within them. Neurons die, neurotransmitters are depleted, and brain tissue is

chronically inflamed. Sufferers slowly lose their memory, judgment, and

knowledge of who and where they are. 14

The course of Alzheimer’s disease varies from patient to patient, but

one frequent pattern interested us. Memory loss is an early and noticeable

symptom of the disease, and it includes memory for words. Patients have

difficulty in retrieving uncommon words, in naming objects, and in

supplying the word that goes with a definition. Yet many patients speak

fluently and grammatically, understand sentences with relatively complex

syntax, and even convert ungrammatical sentences into grammatical

ones.15 The greater impairment in word retrieval than in grammatical

processing may be caused by the distribution of the neurofibrillary tangles

in the cortex. Typically the tangles are more numerous in the temporal lobes

and adjacent parts of the parietal lobes than they are in the frontal lobes, as

shown in this diagram, where darker shades indicate more tangles:16

We predicted that patients with Alzheimer’s disease who have particular

difficulty with word retrieval should look like anomic patients when

producing past-tense forms, and indeed they did. The Alzheimer’s patients

had more difficulty inflecting irregular verbs than regular verbs, were

surprisingly good at the wug-test (84 percent correct), and often

overgeneralized the regular past tense to irregular verbs in childlike errors

such as swimmed (27 percent of the time).17 The psychologists David Balota and Richard Ferraro showed that the same thing happens in reading

aloud: Alzheimer’s patients often regularize irregular spellings, for

example, pronouncing pint as if it rhymed with mint. 18

Is there a contrasting neurodegenerative disease that might dissociate

words from rules in the other direction?

Ullman thought of a possibility. Many neuroscientists believe that the brain has two major memory systems,

one for facts—“knowing that”—and one for skills—“knowing how.” 19 The

fact system, also called declarative memory, needs the hippocampus (a

seahorse-shaped organ embedded in the inner surface of the temporal lobe)

and adjacent structures to form memories; once formed, the memories are

permanently stored in the cortex, largely in the temporal and parietal lobes. These are the parts of the brain that are hit earliest and hardest by

Alzheimer’s disease. The skill system, also called procedural memory,

underlies motor skills such as reaching and walking, but also cognitive and

perceptual skills such as scanning, sorting, ordering, predicting, and

generating associations. The skill system needs the basal ganglia, a set of

organs buried in the cerebrum that receive input from all over the cortex

and send their output primarily to the frontal lobes (via the thalamus, the

relay station in the center of the brain).

Most areas in the frontal lobe have

corresponding areas in the basal ganglia, and these two parts of the brain

work together as parts of a single system. 20

Many neurons in the basal ganglia transmit signals to one another by

releasing the neurotransmitter chemical called dopamine. In Parkinson’s

disease the cells that manufacture dopamine degenerate and the basal

ganglia malfunction. (The most famous young sufferers of Parkinson’s

disease are the boxer Muhammad Ali and the actor Michael J. Fox, who

was diagnosed with the disease in 1998 when he was thirty-seven.) People

with Parkinson’s disease have tremors and difficulty initiating movement,

and when they do budge, their movements are often slow and rigid. They

also may be impaired in the kinds of tests that tap frontal lobe functions,

such as planning, sequencing, and paying close attention. Interestingly, their

speech is often grammatically simplified, with more nouns and verbs and

fewer grammatical morphemes such as prepositions. They have difficulty

understanding sentences via their syntax, such as It was the boy that the girl

tickled and The eagle that the hawk chased was fast.21 Yet their vocabularies often are less impaired and sometimes are not impaired at

all. 22 In some ways Parkinson’s disease is a mirror image of Alzheimer’s. The degeneration affects the skill system rather than the fact system, it has a

bigger effect in the frontal lobes than in the temporal and parietal lobes, and

it compromises grammatical processing more than word lookup. As with all neurological diseases, Parkinson’s patients differ from one

another, and we focused on a sample of patients with slowness in moving

the right side of their bodies. The right side is controlled by the left

hemisphere of the brain, and Ullman figured that these patients were likely

to have more of a malfunction in their left basal ganglia, which in turn

should compromise the language processing areas of the left frontal lobe. As predicted, these patients were slightly better at inflecting irregular verbs

than regular verbs (even when the verbs were equated for ease of

pronunciation, as in passed and lost), were poorer still at inflecting novel

verbs like plam, and never overgeneralized the rule to irregulars in errors

such as swimmed.23 All three outcomes differ from those of the

Alzheimer’s patients, completing a double dissociation that mirrors the one

between agrammatic and anomic patients. In all of these demonstrations damage to a part of the brain causes a

difficulty in retrieving words or applying rules, suggesting that those parts

of the brain may be necessary for those parts of language. But to be sure

that a part of the brain really is linked to some part of the mind,

neuroscientists like to show the opposite as well: that activity in a part of

the brain causes a particular experience or behavior. In the case of a role for

the skill system in carrying out rules of language, a very different

neurological disorder offered the possibility of linking brain activity to

overt behavior. Huntington’s disease is an inherited neurodegenerative disorder made

famous by the folk singer Woody Guthrie, who died of the disease in 1967

(his last days were portrayed in the movie Alice’s Restaurant, written by his

son Arlo).

People who carry the gene begin to notice symptoms in their

forties, when neurons in the basal ganglia start to die. Unlike the

degeneration in Parkinson’s disease, the dying neurons are in circuits that

ordinarily suppress movement, keeping the body under control. As a result,

Huntington’s patients may suffer from involuntary and unsuppressable

movements—hence the old name for the disease, Huntington’s Chorea,

from the Greek word for “dance” that is also seen in choreograph and

terpsichorean.24

Ullman tested several patients with Huntington’s disease and made a

remarkable discovery: They seemed to overapply the past tense rule, as if

the disease led to excess activity in the circuitry that executes mental rules

as well as in the circuitry that executes movement. The patients often

applied the rule to irregular verbs, resulting in errors like digged. Unlike the

similar errors of patients with anomia and Alzheimer’s disease, though, the

errors of the Huntington’s patients could not be attributed to a difficulty in

retrieving dug; they had little difficulty retrieving words in general. The

errors come from a failure to suppress the rule, not from a failure to retrieve

the irregular form. Moreover, the patients often overapplied the suffix to

regular verbs, or applied it too strenuously, resulting in errors such as

lookeded and look d that were rare in all the other patient groups. These

errors weren’t simply stutters or other exaggerations of the physical

movements of the tongue and mouth, because in an error like digged or

dugged the correct form does not call for -ed. Also, the patients never added

an extra suffix to irregular forms that happen to end in t or d, like kept; that is, they virtually never made errors like kepted or kep d. This suggests that

we were observing a compulsion to add the suffix -ed. All this adds up to suggest that irregular and regular inflection, and

words and rules more generally, depend on different systems in the brain. Moreover, if Ullman is right, these two systems may be subdivisions of the

brain’s two major systems for remembering information: Words may be a

part of the “knowing that” system; rules may be a part of the “knowing

how” system. 25

If the 1990s will be remembered as the Decade of the Brain and the dawn of

cognitive neuroscience, the first decade of the 2000s may be remembered as

the Decade of the Gene and the dawn of cognitive genetics. New techniques

for analyzing the human genome are beginning to identify the genes that

shape the brain to learn and feel in particular ways. Two recent discoveries

of genes tied to language and thought will probably be the first of many,

even if they never lead to the scenario in the cartoon on the following page. In the early 1990s Noam Chomsky’s hypothesis that language has a

genetic basis was thrown into the spotlight when newspapers reported the

discovery of a large English family, the KEs, in which half the members

had a congenital difficulty with speech and language.26 The syndrome is

called Specific Language Impairment (SLI), and like most labels slapped on

children with behavioral problems, it means little, only that the problems

with language are not side effects of hearing impairment, autism,

retardation, or some other identifiable condition. Specific Language

Impairment is a family of ailments that strike about 3 percent to 5 percent

of children. They speak late, articulate poorly, and have trouble learning to

read. 27 Everything improves as they get older, but they sometimes struggle

with language all their lives in the way normal people struggle with a

foreign language. They often make errors in speaking, particularly with

grammatical morphemes, as in Carol is cry in the church. Acquaintances

that strike you as chronically tongue-tied and inarticulate may have a

grown-up form of the impairment. Canadian Prime Minister Jean Chrétien,

beloved from coast to coast for mangling the two national languages with

equal proficiency, is a prime suspect.28

ROBOTMAN® by Jim Meddick

ROBOTMAN reprinted by permission of Newspaper Enterprise Association, Inc. Chrétien s son and brother also have language problems, and every

study that has looked at the relatives of people with language impairment

has found that the impairment runs in families.29 In the KE family the inheritance pattern was striking. Among the thirty-one members spanning

three generations, half were impaired, and their distribution in the family

tree would make any geneticist predict that the syndrome is controlled by a

single dominant gene, or a string of genes lying next to each other on a

chromosome. That prediction was stunningly confirmed in 1998 when a

team of geneticists took blood samples from twenty-seven members of the

family and found a small stretch on the long arm of chromosome 7 that

correlated perfectly with having the impairment. 30 They called it SPCH1,

the first genetic region specifically linked to a speech and language

disorder. The region contains several genes whose products are active in the

brain, including a protein that may play a role in the growth and

differentiation of neural pathways, a molecule that makes neurons stick to

other cells, a molecule used as a signal in tissue development, and a kinase,

one of a large family of enzymes that change the function of a protein. (Many kinases are thought to have a role in neural development and

plasticity.) The geneticists don’t yet know whether it is a mutation in one of

these genes, or a deletion of several of them, that causes the disorder. A single gene rarely targets a trait exactly, and SPCH1 is no exception. Its effect is more like a sloppy brain lesion than a surgical excision of a

single organ. The brains of the impaired family members are abnormal in

several areas, particularly the frontal lobe and basal ganglia. On top of their

deficits with language, the impaired children have difficulty carrying out

sequences of mouth or face movements, and they score lower, on average,

than their unimpaired relatives on nonverbal intelligence tests. Yet many of

the impaired family members have intelligence scores in the normal range,

and some test higher than some of their unimpaired relatives. That suggests

that the language impairment is not simply a consequence of an overall

dulling of the brain; instead it appears to be one of several abilities

compromised by the genetic defect. Nor is the language deficit reducible to

the articulatory problems that the impaired members of the family had as

children; they make errors in writing, comprehension, and judging the

grammaticality of sentences, not just in speaking.31

The impaired members of the KE family often omit or misuse

inflections (a common problem among language-impaired people), but their

ability to name objects was not as severely impaired. They should, then,

find regular nouns and verbs harder to inflect than irregular ones, and

should have trouble inflecting novel words in a wug-test. Ullman and the

linguist Myrna Gopnik tested the prediction, as did the psychologist

Faraneh Vargha-Khadem and her colleagues. Novel words indeed were

vexing; some of the impaired members were at a complete loss as to what

to do with them, and most of the others inflected them less than 10 percent

of the time. Contrary to our prediction, however, regular verbs were no

harder on average than irregular verbs; both were fairly hard. Ullman and

Gopnik discovered one reason why regular words were no harder than

irregular words: Some of the family members consciously applied a rule

they had been taught in school. One muttered “add an s” to herself, another

proudly announced that he remembered to use the rule drilled into him by

his teacher. The other members of the family, Ullman and Gopnik conjectured, may

have done passably well with regular verbs for a different reason. Unlike

stroke victims, people with SLI grow up with their impairment and have

opportunities to compensate by using other strategies. They may memorize

regular past-tense forms as if they were irregular, and recall them from

memory when they need them. That would explain why they were baffled

at the wug-test but did better with real regular verbs. This conjecture led to

a successful prediction: The impaired family members should be highly

sensitive to the frequency of a regular past-tense form in the language,

doing well only with the common past-tense forms, unlike unimpaired

people, who inflect rare and common regular verbs with equal ease. 32

Ullman replicated this test with another group of language-impaired

children who offered cleaner scientific tests of theories about language. The

KE family first came to the attention of researchers because of their striking

pattern of inheritance, not because of the details of their impairment. The

psychologist Heather van der Lely has screened many language-impaired

children and selected a few whose impairments are strictly confined to

language, indeed to the grammatical computations underlying language. Fewer than a fifth of children given the label of Specific Language

Impairment meet those criteria. The children in her group are average or

above average in nonverbal intelligence, and they speak clearly and

accurately. Van der Lely found that 78 percent have first degree relatives

with a history of language impairment. Half of their siblings were affected,

brothers and sisters equally, and though often one parent was affected, in no

case were both parents affected. The pattern suggests that “Grammatical

SLI,” as van der Lely calls it, may be caused by a single dominant gene. 33

One boy, AZ, showcases the specificity of Grammatical SLI. His

nonverbal IQ ranges from 119 to 131, putting him in the top 10 percent of

the population. Yet when he was first tested at the age of ten, his ability to

complete sentences, and to understand sentences whose meanings depended

on their syntax (such as The hoy is tickled by the girl) was at the level of a

five-year-old. When speaking, he left out inflections 75 percent of the time,

as in My Dad go to work. He often left out entire phrases, as in The dog was

poking in, meaning poking his head in a jar. And he avoided recursive

sentence structures common in the speech of four-year-olds, such as Can you ask Mum if I can have an ice cream?

AZ’s problems with language were concentrated in grammar. His

vocabulary was below average but not as dramatically so as his

grammatical abilities. And he had no trouble whatsoever reasoning with

words or using language in socially appropriate ways. He was fine in tests

of deductive reasoning; for example, when told that “Mary has never

flown,” he correctly inferred that Mary has never been in a helicopter. He

completed verbal analogies such as “Kipper is to fish as cheddar is to\_.”

And he never made the egocentric error typical of younger children:

opening a conversation with he or she without stopping to think that the

listeners have no idea who he or she refers to. Ullman and van der Lely gave the group of grammatical SLI children,

who were nine to twelve years old, a list of verbs to put in the past tense,

and compared their performance with the performance of control groups of

unimpaired children matched in tests of sentence comprehension (around

five or six years old) and matched in vocabulary (around seven or eight

years old).

Obviously the impaired children would have been trounced by a

control group matched in age; these younger control groups, matched

instead in language abilities, were intended to highlight qualitative

differences in the impaired children’s language, as opposed to mere delays

in their timetable that would have made them like younger children.34

The grammatical SLI children were desperately bad in the wug-tests,

inflecting only about 7 percent of the verbs. They were almost as bad at

inflecting low-frequency regular verbs such as to flap, succeeding 11

percent of the time. The control groups, which were much younger, did

three to seven times better. The impaired children did better with higher-

frequency regular verbs like rob than with lower-frequency regular verbs

like flap, unlike the control children, who were no better with the common

regular verbs than with the rare ones. 35 The impaired children were no better with low-frequency regular verbs than with low-frequency irregular

verbs; the control groups were up to twice as good. Evidently the SLI children were memorizing their regular forms. In an

ingenious follow-up van der Lely found a way to corroborate this

conclusion. Recall that both adults and unimpaired children say that a

monster who eats mice is a mice-eater, but that a monster who eats rats is a

rat-eater, not a rats-eater. That is because mice is a stored root, just like any other simple word, and is available for insertion into a compound,

whereas rats is formed by a rule that creates a complex word later in the

processing stream. Yet the SLI children, unlike the control groups, were

happy to say rats-eater; they said it almost as often as they said mice-eater. That suggests that their mental representations of regular and irregular

forms work the same way.36

All this suggests that the loss of certain genes can interfere with the

development of normal grammatical circuitry in the brain, including the

ability to inflect new and uncommon regular verbs. Children lacking these

genes can learn to compensate by relying more on memory. Are there genetic disorders that go the other way, with preserved language

and impaired intelligence?

That double dissociation would be good

evidence that the human genome codes for a brain in which language is a

distinct system. If language were simply another accomplishment of a

general-purpose intelligence, then any impairment of intelligence would

have to impair language as well. In a recent paper the psychologist Ursula Bellugi and her colleagues

discuss a girl they have worked with for several years:

In describing her future aspirations, Crystal, a 16-year-old

adolescent, states: “You’re looking at a professional book writer. My

books will be filled with drama, action, and excitement. And

everyone will want to read them. I’m going to write books, page

after page, stack after stack...

I’ll start on Monday.” Crystal

describes a meal as “a scrumptious buffet,” an older friend as “quite

elegant,” and her boyfriend as “my sweet petunia”; when asked if

someone could borrow her watch, she replies, “My watch is always

available for service.” Crystal can spontaneously create original

stories—she weaves a tale of a chocolate princess who changes the

sun color to save the chocolate world from melting; she recounts

with detail a dream in which an alien from a different planet

emerges from a television. Her creativity extends to music; she has

composed the lyrics to a love song. In view of her facility with language, proclivity for flowery,

descriptive terms, and professed focus on drama and action, her

aspiration may seem plausible; but in fact Crystal has an IQ of 49,

with an IQ equivalent of 8 years. At the age of 16, she fails all

Piagetian seriation and conservation tasks (milestones normally

attained in the age range of 7 to 9 years); has reading, writing and

math skills comparable to those of a first or second grader;

demonstrates visuospatial abilities of a 5-year-old; and requires a

babysitter for supervision. 37

Crystal has Williams syndrome, a rare form of retardation accompanied

by heart and circulatory defects, an elfin or pixielike face, and abnormal