Learning Distributed Representations of Conceptual Knowledge and their Application to Script-based Story Processing

Abstract

We propose a new method for developing distributed connectionist representations in order to serve as an adequate foundation for constructing and manipulating conceptual knowledge. In our approach, distributed representations of semantic relations (i.e. propositions) are formed by recirculating the hidden layer in two auto-associative recurrent PDP (parallel distributed processing) networks, and our experiments show that the resulting distributed semantic representations (DSRs) have many desirable properties such as automaticity, portability, structure-encoding ability and similarity-based distributed representations. We have constructed a symbolic/connectionist hybrid script-based story processing system DYNASTY (DYNAmic STory understanding sYstem) which incorporates DSR learning and 6 script related processing modules. Each module communicates through a global dictionary, where DSRs are stored, DYNASTY is able to (1) learn similarity-based distributed representations of concepts and events in everyday scriptal experiences, (2) perform script-based causal chain completion inferences according to the acquired sequential knowledge, and (3) perform script role association and retrieval during script application.