Sound advice

For readers of this newspaper, it is a self-evident truism that the word did contains three sounds, and that the first and last of these are the same. But these facts are self-evident only because the reader has mastered an alphabetic script that imposes such an analysis on spoken words. Users of scripts not based on pronunciation, such as Chinese script, have not found it natural to think of words as sequences of consonants and vowels.

Brian Byrne discusses experiments by his group in Australia, mainly involving pre-school children who had not yet learned the alphabet, designed to uncover the structure of the mental processes through which English-speaking children take their first steps in literacy. The experiments seem to demonstrate that preliterate children's default assumption about writing is that it will be of the Chinese type, in which graphic elements represent meaningful units such as words rather than sounds; and also that, even when children have grasped the idea that a script is based on sound, they expect it to represent large phonological units such as syllables, rather than perceiving the individual phonemes of words as elements that can be separately symbolised. To succeed as readers, children have to learn a set of correspondences between written letters and spoken sounds, and also the ability to chop words up into sounds and identify instances of the same sound in different positions. These are shown to be separate cognitive achievements, neither of which suffices on its own, and neither of which occurs as a natural response to simple exposure to data.

The significance of Byrne's findings is that they amount to a strong objection to techniques of teaching reading, such as "look-and-say'', which assume children can induce what they need to know about alphabetic writing from examples. Byrne argues particularly against K. S. Goodman's "whole language'' theory, according to which learning to read and write is as natural as learning to speak, requiring only motivation.

Byrne is by no means the only researcher who has been coming up with findings pointing in broadly similar directions, and his book contains appropriate references to others' work. But his experiments contain novel features, and they fit together well as a self-contained demonstration of each step in the argument.

Walter Kintsch aims to draw together a programme of research carried out over the past two or three decades into a global framework for understanding human cognition. Kintsch calls his framework the "construction-integration'' theory, because it models comprehension as a two-phase phenomenon: a person initially responds to input (such as reading a text) by building an incoherent mental structure based purely on the contents of the input, and then achieves coherence through a constraint-satisfaction process of activation spreading through a wider cognitive network. Kintsch ex£ this theory in his first four chapters, and in the remaining seven chapters he discusses its application to specific examples. Kintsch's book is offered to his fellow academic psychologists as a definitive statement of his distinctive cognitive paradigm.

Some of Kintsch's discussion of concrete application areas contains fascinating novel findings. His chapter on differential reading skills cites recent experiments by his group at the University of Colorado that seem to demonstrate that good students learn less from texts in which intellectual links between successive items of information are carefully spelled out, than they do from prose that forces them to work at inferring the links for themselves. In these days when academics are subordinated to managers who demand that teaching should be made accessible to ever-weaker students, it is good to have hard evidence of what is lost by such policies.

On the other hand, the case for Kintsch's novel "architecture of cognition'' seems less persuasive. The concept of building cognitive structures and integrating them into a knowledge network by minimising energy is not particularly controversial; many researchers internationally would agree that it must probably be something like that. But one has not really got a concrete paradigm until the model is made much more specific.

What, for instance, is the nature of the structures built in the "construction'' phase? Kintsch has two quite different answers. Sometimes he says that the reader of a text responds to it by building a labelled tree structure, broadly akin to the formulae of logicians' predicate calculus or linguists' semantic parsing - though Kintsch claims that, for his purposes, it is not necessary to deal with subtle logical considerations. At other times, Kintsch discusses the construction phase in terms of the latent semantic analysis theory developed by his colleague Thomas Landauer. This theory uses statistics of word co-occurrences in a large knowledge base such as an encyclopedia, together with ingenious mathematical manipulations akin to factor analysis, to represent a word, or a sentence, as a point in a space of several hundred dimensions. Kintsch repeatedly asserts that these alternative methods of representation are loosely equivalent. But, on the face of it, they seem unrelated. And, while Landauer's analytic technique may have useful applications, it is hard to see how a point in a multi-dimensional space is the sort of thing that could be equated with the meaning of a linguistic form.

Kintsch recognises that sceptics may find the construction-integration model too vague to count as a testable theory. He says that the theory makes strong claims; but not all readers will be convinced. The construction-integration examples he displays to illustrate his later chapters look like toy models tailored to the empirical data; if the observations had been different, the models could readily have been changed to suit them, so that the theory is something of a fifth wheel.

One problem is that Kintsch pays little attention to the many programmes of research being conducted within frameworks not unlike his in places other than Colorado. He repeatedly laments the fact that there exists no draft of a representation of general knowledge in the form of a logical network; what he wants sounds rather like Douglas Lenat's CYC system, but Lenat is never mentioned. The book is quite unusual in the proportion of its frequent literature citations which are self-citations; even when items by other authors are quoted, these often turn out to be Kintsch's graduate students. Scholarship succeeds best when it is more outward-looking than this.

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