Natural Language Interface using Contextual Information by Connectionist Model

Abstract

In natural language processing, context dependency plays an important role. However, traditional symbol processing methods, based on static structure and lacking flexibility, make it difficult to process contextual information. Connectionist models offer an alternative method for context processing which provides dynamic structure and flexibility. For these reasons, many researchers have applied the connectionist model to problems in natural language processing, including that of context processing. For example is the DISCERN system which is based on script structure built from several neural networks. This system reads stories and answers questions about them. But this system can not process contextual information except script-based information. This paper describes our context-exploiting natural language interface system. This system uses recurrent neural network for representation of the user model. This represents the context dependent relations between a user's natural language requests and his intended application-system actions. Through learning in the recurrent network, this system can interpret the user's intended meaning of context-dependent sentences. Its context processing ability differs from that of DISCERN in many respects. For example, our system can interpret pronouns and automatically acquire the meaning of words.