Book Reviews
Words in the Mind: An Introduction to the Mental Lexicon Jean Basil ISBN ISBN
Aitchison Blackwell, Oxford, 1987. x+229 pp. O-631-14441-2. L25 (hb) 0-631-14442-O. 0.95 (pb)
Danny JONES Centre for Computational
Linguistics,
UMST, Manchester
The author’s main objective is made clear at an early stage: “The overall aim of this book is to build a ‘model’ of the mental lexicon” (p. 34), and in order to accomplish this goal, she explores the issues of how the brain acquires, stores, uses, recognises and generally manipulates words. Thus the scope of the work is broad, but due to its introductory function it does not examine any one particular aspect of psycho-lexicography in any great depth at the expense of any other. The book is well-structured and balanced in this respect. No prior detailed knowledge of psycholinguistic practices or issues is assumed and the author’s style is far from being pedagogically intimidating. In fact the complete opposite is the case and it is this which is initially one of the most striking aspects of the book: literary references and drawings are to be found in abundance illuminating and facilitating the discussion of difficult concepts. However, there is always a danger of becoming just a little too verbose when wanting to avoid overly technical content and didactic style. While this is an admirable consideration for readers with very little experience of a given field of study, anyone with more knowledge may find the text just a little too circuitous in places. The constant use of literary references for exemplification may be off-putting to some but welcomed by others. The author may indeed be giving the reader a subtle reminder of the importance of not losing sight of how language is actually used rather than solely relying on invented examples. Unfortunately it is impossible to be all things to all persons but it would be a
Machine
Translation
4~4 (I989)
BOOK REVIEWS shame if anyone were put off using the book as a general source of reference even though a little patience may have to be exercised in order to arrive at a desired goal. Although the author’s main aspiration is the construction of the human lexicon, she also points out, however, that models are often metaphorically inspired simplifications of the real thing, e.g., libraries or computers. There may be a big difference between an abstract model which is built within the confines of a book and the somewhat more rigorous demands imposed on a model encoded in a computer program. This is the methodological interface where the software of the brain and the software of the computer meet and it is on this ground that the author details some of the most interesting contemporary psycholinguistic models. According to the theories described, the human brain appears to organize words in networks (semantic fields) with some connections stronger than others: co-ordination and collocation relationships are the strongest links which hold words in relation to one another (p. 85). But what relevance has information like this to the computer processing of natural language? Let us look at some of the theories in slightly more detail. The author describes three different models of mental dictionary look-up for language production: a stepping-stone model, a waterfall model and an electricity model. The tirst assumes that a word and its associated phonological realization are linked in a one-to-one serial relationship. Such a model says that the word-meaning choice is made and then we move to the second phonological stage. Any errors which occur are due to independent selection errors in either stage. Such a model, however, fails to account for speech errors where semantically related words are chosen by mistake, e.g., Don’t contact lenses make your ears [eyes] sore? @. 171). Such errors can only be accounted for if a person is still processing meaning possibilities whilst processing in the second phase. The cascade or waterfall model proposes a concurrent solution to this problem whereby multiple lexical possibilities are accessed at the same time and these remain available for selection as the phonological stage is entered. The overgeneration of word choices which would result from this kind of process can indeed account for errors like the above as lexical information would still be available even after dictionary look-up. The electricity model is cited as a means of allowing information to feed back into the cascade. Spreading activation - a process whereby activated elements themselves activate neighboring and related elements - would therefore help to narrow the possibilities until the appropriate word is located. Speech recognition also appears to employ concurrent dictionary look-up techniques. Recognition is assumed to be modeled most accurately by assuming the facility of spreading activation where multiple possibilities are taken from the lexicon based on the very beginning of a word which is heard; and as more of the word is processed, activations toward semantic interpretations are strengthened in some cases and weakened with others. Backtracking occurs when incorrect
WORDS IN THE MIND
301
guesses are discovered. Feedback from the semantics also influences activations (as with production, but with the direction of feedback reversed). It is therefore clear that human dictionary access occurs in parallel. It is only when we assume concurrent models that many phenomena of language use can be adequately explained. An inescapable truth of natural language processing is that it does not matter how powerful the analysis tools are in any given system: if the machine comes across a word it does not recognize, everything comes to an abrupt halt. Although techniques can be employed to allow processing to continue, it is obviously preferable to have as full and complete a lexicon as possible. A great many systems have been designed to work within the domain of a sublanguage or restricted domain, e.g., the MT system METEO. Such a facility obviates the need for a full-scale lexicon, which is what would be required if restrictions on input were not imposed. However, if more generally usable systems are going to be designed, dictionary size is going to have to be very much increased. Given the fact that the average human brain can store up to 250,000 words, perhaps different approaches to dictionary look-up and analysis may have to be adopted for large-scale computer implementation. If cognitive modeling is chosen as the driving force behind computational linguistic models of language, then much of the ensuing formal description will have to be expressed with the ultimate aim of parallel execution, as conventional serial execution would not only be too slow but not a proper reflection of the formalism of which runable programs would be the embodiment. It makes sense to do dictionary look-up in parallel. Not only does it reduce overall processing time, but it has implications for general system design and methodology. Indeed, parallel programming languages themselves are becoming more widespread. Parlog, for example, is a concurrent version of Prolog and thus is highly suited to language processing even though parallel hardware is not available (concurrency is simulated by the compiler). Thus we have a situation in which “a degree of abstraction becomes possible in which the order of doing things becomes a detail that programmers can choose to ignore” (Conlon, 1989:ix). Obviously this is a desirable aim for any cognitive-based language theory as it would appear from the author’s descriptions of the mental lexicon that concurrency is the thread which runs through all lexical processing facilities. Adopting a concurrent methodology has a number of advantages within the domain of the dictionary. It would allow multiple word categories to be extracted at one procedure call. If one category fails then the alternatives are already available for use without having to go back into the lexicon. However, the advantages of using parallelism need not stop at the level of the word. The syntax and semantics of language need to communicate with one another. Lytinen (1984:4f) says that there are five principles regarding the relationship and integration of syntax and semantics, the first being that syntactic and semantic processing should proceed at the same time, although the main argument against this approach is that the concept of modularity becomes endangered. However,
BOOK REVIEWS this need not be the case at all as syntax and semantics can remain separate but work at the same time with the process of spreading activation mediating between the two processes. NLP theories and their implementation may be hampered by the realization that with conventional serially-based software and hardware, it may not be practical to employ techniques which might involve large-scale and potentially uncontrollable search spaces (spreading-activation on a serial machine could, potentially, be very expensive). If large-scale concurrency were available, models of natural language analysis, generation, and translation could perhaps benefit from the techniques employed so effectively by the brain itself. REFERENCES Conlon, T. 1989. Programming
in Parlog.
Wokingham:
Addison-Wesley.
Lytinen, S.L. 1984. The Organization of Knowledge in a Multilingual, Integrated Parser. Doctoral dissertation, Department of Computer Science, Yale University, November 1984.