A neural network model for acquisition of semantic structures

Research suggests that natural language processing (NLP) can be profitably viewed in terms of the spread of activation through a neural network. However, since the critique by Fodor (1988) of the style of connectionist representations, one of the biggest challenges facing proponents of connectionist models of NLP is the rich structures of language. As models of NLP, neural network systems must exhibit the properties of compositionality and structure sensitivity. This paper describes a neural network model in which simple recurrent network and recursive auto-association memory are combined to acquire the semantic structures from sentence constituents. This imposes no prior limit on sentence structures. The model can be viewed as a tool of conceptual acquisition and generalization extraction in language understanding