With Words and Rules Steven Pinker returns to the topic of language, a domain quite fertile for him in the past. In his first book, The Language Instinct, Pinker inferred from the structure of language facts about its evolution. His newest work uses structure and evolution as a starting point to construct a theory of how language is produced and interpreted within a brain. As the history of a single brick may tell us much about the edifice in which it occurs, observations on just a single part of speech can be the basis for hypotheses about language in general. Pinker adopts this minimal approach, never straying far from the rather specific topic of conjugation of irregular verbs. "[I]t is easier to compare two great big theories," he writes, "when each is vested in a highly specific hypothesis and the hypotheses compete on the same ground." What marches across this narrow ground, however, may surprise. Along the way to answering the question of irregular verbs, Pinker tramps through rationalist and empiricist epistemologies, symbolic and connectionist models of computation, and cognitive and behaviorist psychologies. The truth about all these dualities will be illuminated, he says, if we just think hard enough about irregular verbs. This is a mighty challenge, but Pinker has an arsenal of facts with which to tackle it.

Three and a half centuries ago, the rationalist philosopher Renee Descartes decried the haphazard layout of science and philosophy. He compared it to the plan of an old European city, grown from an ancient settlement hemmed in by twisting streets and walls. The acceptance of such historical legacies struck Descartes as fundamentally unreasonable. Belief, he wrote, should be a well organized system, proceeding only from orderly inferences founded on innate and self-evident axioms. In the modern field of linguistics, Descartes' rational derivation of knowledge is mirrored in Noam Chomsky's idea of a universal grammar, a system of rules for acquiring and using language that is innate and `hard-wired' into the human brain.

Pinker reviews Chomsky's ideas of language organization, and with a series of engaging, everyday examples reveals that the human faculty for language does indeed have the modular design of a well ordered system. To separate semantics (the meaning of a word) from morphology (the form of a word), and thus to confirm de Saussure's observation of the arbitrariness of linguistic representation, Pinker cites words such as `become', whose meaning bears little relation to that of the root `come' but which is nevertheless conjugated similarly. To separate morphology from phonology (the low-level representation of words as sequences of sounds), he cites the example of mispronunciations such as `froos' from `fruits' where the omission of the morpheme represented by `t' causes a predictable, phonologically based change in the voicing of the `s'.

These successes bode well for a linguistic theory based solely on rules. But there are some gaps. If all verbs are conjugated according to morphological rules, Pinker asks, then why are irregular verbs so much less well behaved than regular verbs? People's memories for, and usages of, irregular past tenses can be completely separate from memories of the root words -- Pinker cites `wrought,' which actually is a conjugation of an archaic form of the verb `to work.' When response times are precisely measured in laboratory experiments, the speed with which people form conjugations depends on word frequency for irregular verbs, but is constant for regular verbs. And irregular past tenses call to mind the corresponding present tenses less potently than do the past tenses of regular verbs. All these findings suggest that irregular forms are retrieved from memory rather than computed by application of morphological rules. What alternative, Pinker wonders, might there be to linguistic computation based solely on rules?

Set against the rationalist philosophers were the empiricists, who held that the sole source of knowledge is sensory experience, that human theories may be consistent with observations but are fundamentally unverifiable and have no validity independent of those observations. Empiricist thinking had a great influence on modern psychology in the form of the behaviorism of B.F. Skinner. According to Skinner's premise, the only way of understanding an organism was to observe the conjunction of stimuli and responses, and to make inferences about how inputs and outputs are related. The way to understand language, then, would be simply to connect the words a child says with the words a child hears. This empiricist strategy sounds as if it may explain the apparent involvement of memory in the conjugation of irregular verbs.

Of course, a child learning to talk is rather different from a psychologist learning to theorise. The child is unaware of the tortuous history of the language, and seeks not an explanation for grammar, but a way of producing grammar. Put another way, the interest of the philosopher or the psychologist is descriptive, whereas the interest of the child is operational. Unless Pinker can cast the rationalist-empiricist debate in more operational terms, his argument will miss the point entirely. Fortunately, Pinker anticipates this conundrum. His operational argument comes in the form of a contrast between symbolic and connectionist models of computation.

Symbolic processing is the bread and butter of artificial intelligence, and, so its skeptics claim, the reason that artificial intelligence has never been able to mimic the generality and flexibility of biological intelligence. From a symbolic point of view, ascertaining the sense of a word purely by morphological rules would be rather like the procedure a computer processor goes through in decoding an instruction. In the computer, an instruction code carries no redundancy; an error in a single bit throws off the entire computation. Pinker cites the nineteenth-century example of Bishop John Wilkins's invented `philosophical language,' which took morphological representation to an absurd extreme. In Wilkins's language every morpheme, every bit of every word, represented some category within which a representation was deemed to reside. The very compact logic of such a system, says Pinker, not only makes it prone to errors but also leaves no room for the addition of new terms. Human languages, he observes, don't work this way.

The alternative to rationalist symbolic processing, in which every operation is the application of a rule, is empiricist connectionism, in which every operation is a generalization from remembered examples. Connectionism employs networks of small computing elements, modeled loosely on the networks of neurons that make up brains. A connectionist network begins as the empiricist's tabula rasa, able only to guess at the correct outputs for the inputs fed to it. As the network receives feedback on its errors, it gradually learns to produce correct outputs. Where the symbolic system relies on computational power to instantiate from rules, the connectionist system relies on memory capacity to generalize from previous experiences.

Against Chomsky's rationalist notion of innate linguistic rules, then, comes an empiricist model in which verb conjugations (not to mention other morphological transformations) are based on a child's experience and memory of the correct forms. Here at last, we are led to believe, is a true explanation of why people seem to memorize irregular forms even when those forms could be explained by morphological rules. If the rule-based approach has some holes, though, the connectionist one fails in complementary ways.

Pinker is at his best when describing and untangling the pressures that lead to language's gradual transformation. He entertains with a series of etymological detective stories on topics such as why we say `seeped' but not `keeped,' how `snuck' snuck into the language, and why baseball fans seem to have a language of their own. Along the way he cites many more examples of forgotten rules, digging up the foundations of Descartes' old city. Indeed, his metaphors borrow from geology; he writes, for instance, of the "erosion of the distinction between participles and past tense forms," an ongoing transformation in which syntax has taken up the functions formerly handled by morphology. In Old English, at the other extreme of this transformation, much of the meaning of a sentence could be derived from the declensions of its individual words. Readers who enjoy this sort of etymological sleuthing will be richly rewarded by these descriptions alone; their use in explaining how the brain processes language is a bonus.

After giving his synthesis of empiricism and rationalism in the words that we reproduce and the rules that we learn to govern them, Pinker steps back from neurolinguistics and considers his theory's implications for cognition. He examines the failures of models of cognition based solely on rationalist or empiricist principles, and suggests, as he did in How the Mind Works, that the evolution of language may explain a great deal about the evolution of human cognition. It is in these final speculations that Pinker seems on less solid ground: it's fine to instantiate from the larger epistemological debate to the relatively confined field of verb tenses, but generalizing in the other direction seems a leap of faith likely to draw skepticism from rationalist and empiricist alike. It's difficult not to forgive Pinker for taking such liberties, though, given the informative and entertaining results of his previous excursions.