Language, Connectionist and Symbolic Representations of

Connectionist or neurocomputational representations are based on sets of similar, multiple connected neuron-like parallel processing units, the connections bearing modifiable weights adjusted on the basis of locally available information using a learning algorithm. Symbolic representations, by contrast, are defined in terms of rules relating expressions in a formal language. Among the claims that have been made for connectionist models is the ‘emergence’ of generalizations that had been thought to require the mediation of rule-based grammars and modular symbolic processing architectures. A more convincing linguistic role for such networks lies in their potential for inducing grounded conceptual structure and statistical models as infrastructure for acquisition and processing of standard symbolist representations of lexicalized syntax and semantics.