Irregular link-ups, regular rules OK

Geoffrey Sampson enters the complex world of connectionism and irregular verbs, only to find an entertaining work that lacks a wider purpose.

After one superbly written bestseller ( The Language Instinct ) about the nature of human language and another well-received book ( How the Mind Works ) on the workings of cognition, Steven Pinker has followed up with a third book, again aimed at a general readership, whose topic is irregular verbs (together with some material on irregular noun plurals).

On the face of it this might look like something of an anti-climax. The key to understanding why Pinker sees the subject as comparable in significance to those of his earlier books lies in a development that occurred in the cognitive sciences in the 1980s. Until that time, it was more or less universally believed that children master the grammar of their language by inducing general rules from examples uttered by their elders. The adult's knowledge of his mother tongue is encoded in his mind as a structure of rules, supplemented with information about individual exceptional cases. In the 1980s, this consensus was challenged by a computer-based model of cognition called "connectionism" or "parallel distributed processing", which generated regular behaviour from a knowledge structure that contained nothing resembling general rules. A connectionist model is a heap of many simple units linked in simple or random patterns, reminiscent of the neural anatomy of the brain. Some units ("input" and "output") correspond to simple components of perception and action, for instance phonemes, and rule-like behaviour is induced in the model through repeated exposure to examples. This gradually strengthens or weakens individual connections, bringing about changes in the overall propensity to associate complex percepts with complex behaviours.

It happens that one of the most impressive early connectionist successes was research by David Rumelhart and James McClelland, which produced a connectionist simulation of children's acquisition of English past-tense inflexion. The model did a reasonable job of reproducing the observable behaviour of real children at different stages of language learning, yet even when it had mastered regular past tenses, it contained no identifiable element corresponding to a "past-tense rule".

Rumelhart and McClelland's initial model worked well but, of course, not perfectly. Pinker was a leader of the resistance to connectionism: he and his colleagues executed a series of psycholinguistic experiments designed to establish that human beings really do induce general rules from linguistic experience. Part of the aim of Words and Rules is to popularise this line of research. In fact, Pinker adopts a compromise position: he now believes that the connectionists may be broadly correct about the mental processing of irregularities in language, while the rule-based orthodoxy is correct about the processing of the regular cases.

Pinker's effort to establish the truth of this position is certainly ingenious, and persuasive, in the way that it anticipates various possible alternative explanations for aspects of language behaviour and comes up with crucial experiments that refute them. As in his earlier books, Pinker throws out remarks that deserve to become classic linguistic aphorisms, for instance: "Those who cannot remember the past are condemned to compute it" to express the alternatives of case-based and rule-based techniques for generating correct inflected forms. Along the way, too, Pinker offers readers who may be more interested in the English language than in hypotheses about psychological language-processing machinery what must surely be the most complete popular account available anywhere of the facts about English irregular verbs. It is interesting, for instance, to learn that several irregular verbs (among them "knelt" and "caught") came into popular usage as late as the 19th century, but that no new irregular verb has appeared in the 20th century (the newest irregular, Pinker tells us, is "snuck" as an informal variant of "sneaked", recorded from the 1880s onwards).

I was rather sorry to find that Pinker omits my favourite English irregular form, "tret" as the past tense of "treat" (perhaps because it is not found in North America). "Tret" has never been accepted as part of standard metropolitan English, but it is long established (dating at least from the 16th century), and in the Northeast it is routinely used by people who would not normally resort to dialect or slang usage; sometimes one sees it creeping past editors into print. "Tret" is a very unusual irregular verb, because it is the only case in British English (apart from past tenses that are identical to present tenses, such as "cost") where the root does not belong to the native Germanic vocabulary. ("Treat" is cognate with French " traiter " and ultimately with Latin " tractare ". American English has a different example, "pled" as past tense of "plead".)

Despite the general linguistic interest of Pinker's exposition, though, there is a question mark about the wider purpose of this book. Pinker deploys a wide-ranging and ingenious set of arguments and experiments to defeat the connectionist claim that everything in the mind is merely a pattern of weights on connections between simple processing units. But this is worthwhile only if readers come to the book believing that the claim might be true. The average general reader is unlikely, surely, to have a view one way or the other and, within academic cognitive science, Pinker may be close to flogging a dead horse.

Rumelhart and McClelland's original work in the mid-1980s held out an exciting prospect of an entirely novel model of cognitive functioning that looked plausible and appealing to many commentators, myself included. But, as the years went by, that early promise never seemed to be fulfilled. Gaps and problems in the initial model were forgivable, but later attempts to deal with them did not seem very convincing. At the end of the century, there is still a group of researchers trying to make the connectionist manifesto work for language (and other areas of behaviour), but it seems fair to say that they rank now as a sect rather than as the large-scale movement they once seemed destined to become. Pinker discusses the shortcomings of various connectionist attempts to solve the problems of their early models. His arguments are rather convincing, but the more successful they are, paradoxically, the more one wonders whether this book was really necessary.

In the past few years, the idea of linguistic knowledge without linguistic rules has again become a live issue, through work by Rens Bod and others on what they call "data-oriented parsing". But this is quite different from connectionism and Pinker does not mention it at all, although currently it might seem a stronger threat to his preferred cognitive model.

There is a second thread running through the book: Pinker links the defeat of connectionism with support for his theory, advocated earlier in The Language Instinct , of genetically inherited knowledge of language structure. It is true that connectionism implies an empiricist view, in which very few aspects of cognitive functioning are innate. The converse does not hold, though. The idea that an adult's knowledge of his language is held in his mind largely in the form of general rules, is neutral between empiricist and nativist views about how the rules got there. Rules might be encoded in our genes, or formulated by individuals in response to their experience. Pinker uses some passages of Words and Rules to respond to critics of his earlier arguments in The Language Instinct , but these passages are not always as persuasive as other parts of the new book.

For instance, at one stage Pinker responds to a point I made about "headless" compound nouns in the book, Educating Eve , which I wrote as a response to The Language Instinct . Pinker claimed there that when a compound refers to something other than the thing its main word refers to (as "sabre-tooth" refers to a kind of tiger rather than a kind of tooth), the compound will form its plural regularly even if the main root is irregular: two sabre-tooths, not two sabre-teeth. By a chain of argument that I shall not pursue here, Pinker used this alleged generalisation to support his nativist theory of language acquisition.

In Educating Eve I urged that the generalisation was simply wrong. In real-life data, I found plural Blackfoot Indians being called "Blackfeet", and plural pinkfoot geese being called "pinkfeet". Pinker (who does not directly quote my book, but presumably has my argument in mind, since he uses exactly the examples that I introduced into the debate) replies that people who talk about Blackfeet or pinkfeet may actually be thinking of the people or birds as feet rather than as complete organisms. To me this sounds absurd, or circular.

Again, Pinker points out that an individual person may figuratively be called a "silly goose" or a "filthy louse", but he claims that it is "strange" to use the irregular plurals figuratively for groups of people - "silly geese!" or "filthy lice!" Is it strange? I believe someone in the 1970s called his political enemies "lice on the locks of humanity", and Anthony Burgess certainly once called disc jockeys "electronic lice" in the magazine Punch . I am not going to peruse the complete works of Angela Brazil to find an example, but it would not seem to me at all odd or unexpected to find a gaggle of her schoolgirl characters being called "silly geese" by a classmate. Pinker's generalisation just does not ring true.

Furthermore, Pinker concedes that his generalisation about English does not work at all for German, where, for instance, more than one Grömaul ("big-mouth", ie, braggart) are Grömäuler , with the usual irregular plural of Maul . Even if the English generalisation were correct, it could be evidence for a genetically inherited "language instinct" only if it applied to all languages alike. A generalisation that applies just to one language must be a learned aspect of culture, not an inherited biological feature.

So I remain unpersuaded that human beings inherit a language instinct. But Pinker's arguments for the reality of linguistic rules are convincing; and his writing, as always, is great fun to read.

Geoffrey Sampson is professor in computer science and artificial intelligence, University of Sussex.