1.2.4 The Cognitive Linguistics Part

The second practical motivation for CLASPnet has been the desire to develop a connectionist model which can be used for testing some of the tenets of Cognitive Linguistics. I am of course not the first to try to integrate connectionist and cognitive linguistic lines of research: for example, Harris (1990) and Bernsen (1993) have both presented network models which draw on both traditions. What's more, both Lakoff (1988a, 1988b) and Langacker (1990a, 1991) have commented favorably upon the possible benefits of a synergistic effort: calling himself "a cognitive scientist and a linguist" (1988a: 302), Lakoff has claimed "The point of all this is that, counter to what critics like Fodor, Pylyshyn, Pinker, and Prince have said, it is connectionism, not the symbolic [i.e. classical -- EvE] paradigm that is the only game in town" (1988b: 39-40). More recently, Goldberg (1995) has also drawn links between Construction Grammar, Cognitive Linguistics, language acquisition and connectionist modeling. It must be said, however, that all the wooing notwithstanding, few working networks have been presented.

There are three areas in which the implementation of CLASPnet has been influenced heavily by cognitive linguistic ideas:

• Bipolar Input: At the input level, both semantic and formal information are presented simultaneously. Following Langacker (1987), I take language to be symbolic in nature, with each symbol having both a semantic and a formal pole. Consequently, even the so-called function words have been given a semantic representation.

• Perceptual Semantics: The semantic representation which is used to represent the words in the lexicon is mainly based on the perceptual properties of the concepts linked to these words. Hence, abstract concepts have a poorer representation than the so-called content words. Following Lakoff (1987), however, the network is also given the opportunity to develop its own complex representations for these abstract words during learning.

• Emergent Syntax: The main task of the network is to link information at the local (word) level with properties of higher-level constructs like clauses and sentences. By choosing this (somewhat counter-intuitive) task, I want to show my belief in the cognitive linguistic notion that these higher-level -- 'syntactic' -- properties emerge spontaneously from the patterns which are present at the local level.

A final area in which CLASPnet may be considered compatible with Cognitive Linguistics is its use of a functional classification of clause types at the output layer. In this respect, the model is also inspired by work in Construction Grammar (Fillmore 1988) and, to a lesser extent, the various functionalist schools.

Despite my good intentions, CLASPnet is definitely still a far cry from a correctly implemented Cognitive Linguistics theory. This is partly because of the current limitations of computational power (the brain is still a far more powerful 'machine' than the best computer), partly because of the difficulty in implementing cognitive linguistic notions (even 'trajector' and 'landmark' are not straightforward), but mainly, of course, my own fault.