1.2.3 Bridging Connectionism and Theoretical Linguistics

A more practical factor in the development of CLASPnet has been the aim to construct a model which would be capable of arousing the interest of (some) theoretical linguists. This is not meant to suggest that there have not yet been connectionist models of many aspect of language -- even a very incomplete survey like the following proves that idea to be incorrect (see Van Everbroeck 1994 for summaries of these and other articles):

• phonology: Sejnowski & Rosenberg's (1986) NETtalk; Hinton, Plaut & Shallice's (1993) model of deep dyslexia; Dell, Juliano & Govindjee's (1993) model of phonological errors;
• morphology: Gasser's (1993) model of affix acquisition and production, and MacWhinney's (1989) model of German case acquisition;
• syntax: Elman's (1992, 1993) models of complex sentence processing;
• semantics: Gasser & Smith's (1993) model of the conceptual categorization underlying nouns and adjectives; Harris's (1990, 1994) model of prepositions;
• pragmatics: Katz & Dorfman's (1992) model for capturing the coherence of conversations.

None of these models, however, has had much impact in theoretical linguistics proper. One of the reasons for this is undoubtedly that none of the articles mentioned above have appeared in linguistics journals. But conversations with other linguists have made it clear to me that, while they express interest in the field, they would like to see accessible introductory texts. Moreover, those linguists who had actually taken the trouble to read connectionist papers dealing with language mentioned that none of the models seemed very plausible from a linguist's point of view. What usually struck them as especially problematic was the oversimplification of the language input which the net had to process: sentences like 'boy who chases boy chases boy' or 'school-girl stirred kool-aid with spoon' to them did not seem close enough to normal English sentences. Similarly, the small number of words and constructions used in these models fails to give linguists much confidence in their relevance to theories of human language.

Although still based on but a limited subset of written English CLASPnet does try to address some of the linguists' concerns: the context-free grammar used for generating the training and testing corpora was more than 800 lines long instead of the more usual 20 or 50; similarly, the vocabulary used was closer to 300 than to 40. As a result, the sentences which the network had to process exhibit a lot more of the variety of real sentences than the ones used in earlier models. The effects of the combination of formal (i.e. orthographic) and semantic information at the input level should be of interest to linguists who are interested in the relation between the two. At the output level too, CLASPnet tries to be more acceptable for linguists: some of the units used stand for well-known theoretical concepts like voice, polarity, and mood, while others are concerned with different (functional) types of clauses. In this way, I hope that CLASPnet can be a first step at bridging the current gap between connectionist models and theoretical linguistic research.