When a rabbit is a red herring

Nobody knows how children learn the meanings of words. Any adequate theory needs to cover a range of phenomena that have been ignored in this book and in most of the literature as well." Was that a disparaging summary of this book? No, those are Paul Bloom's own words in his concluding chapter.

Bloom, a professor of psychology at Yale University, believes that learning words is an immensely complex affair, far more complex than using them creatively. This belief is such that he does not hesitate in his opening chapter to compare word learning to playing chess and the creative use of words to playing checkers. This said, he immediately turns to the philosopher Willard Quine to substantiate this startling premise, upon which this whole book rests.

Quine asks us to imagine a linguist trying to learn an unknown native language. A rabbit scurries by and the informant says "gavagai". Quine then argues that there is "an infinity of logically possible meanings for gavagai". Indeed, in the circumstances, gavagai could mean "rabbit", or "it is running away", or it could refer to what this particular rabbit is fleeing from, be it a fox, a dog, or a hunter called Gavagai. The rabbit might even have merely happened along as the native was saying "nice day if it does not rain". And yet, somehow, the linguist eventually manages to learn what gavagai means. An all-too-familiar puzzle - I was once in the same position as Quine's linguist, only with mangy dogs and flying foxes in lieu of rabbits. "These problems of reference and generalisation," Bloom comments, "are solved so easily by children and adults that it takes philosophers like Quine and Goodman to even notice that they exist". And it takes linguists who have dabbled in fieldwork and computational linguistics to realise that Quine's rabbit is a red herring.

In a perversely literal way, the introduction on the front flap of the dust-jacket is very true: "This book requires no background in psychology or linguistics." When much later Bloom writes that "some languages, such as Maltese and Chinese, use different words for counting one, two or three items than they use for describing them", I cannot imagine what he might possibly mean. Certainly, numerals behave differently from adjectives in Chinese and in Maltese. But so do they in every language I can think of. But back to Quine's rabbit.

There is an infinity of logically valid interpretations for Quine's rabbit because only one single event involving a rabbit is considered. The fact is well known to epigraphers and cryptologists: given a short enough unknown text, you can read just about anything into it. John Chadwick, who collaborated with Michael Ventris on the decipherment of Linear B, demonstrated how the text of the Phaistos disc - a clay tablet from ancient Crete, c. 1700 BC - was far too short to allow a decipherment: with his tongue firmly in his cheek, Chadwick read there a legal document on wells and water rights, all in good, grammatical... medieval Japanese! But as the size of the corpus grows, the range of tenable interpretations shrinks. The meaning of Quine's gavagai can even be inferred without reference to an obligingly scurrying rabbit.

In his 1993 PhD thesis, Finding Structure in Language , Steven Finch reports how he generated a corpus of English sentences, onto which he applied a variety of clustering algorithms. The dendrograms produced showed such words as dog, cat and mouse in one cluster, girl, boy, woman and man in another. A surprising result, seeing that no further information was fed to the algorithms: no thesaurus, no semantic categories, just the raw corpus - strings of words, spaces and end-of-sentence markers. I have experimented with such algorithms and they have proved computationally inexpensive, negligible compared to Deep Blue, the chess-playing program. Even more surprisingly, significant results were obtained from comparatively small corpora. From a collection of 1,000 answers to the question "How would you define a weed?" provided by a botanist colleague, the dendrograms computed showed such words as "obnoxious" and "unwanted" forming one cluster, and "place", "garden" and "backyard" forming another. There was even an obvious correlation between the sex and ethnic origin of the respondents and the words used. Women spoke of weeds as plants, men as things, and Pakistanis always made reference to them growing in water. The total corpus being about 5,000 words long, the familiar Chomskyan argument of the "poverty of the stimulus" does not apply here.

As Bloom wrote in his conclusion: "Any adequate theory needs to cover a range of phenomena that have been ignored in this book". Ignoring the statistical phenomena uncovered by computational linguistics is perhaps not such a good idea.

Jacques B. M. Guy holds a PhD in linguistics from the Australian National University.