Implementation of sentence meaning with attractor sequences

Hermann Moisl

Abstract

This paper is concerned with engineering of natural language understanding systems. More specifically, it describes a finite state computational architecture for semantic interpretation of natural language strings in which the state sequences that compute the NLU function are inferred from environmental input rather than explicitly designed. The FSA architecture is implemented by integration of attractor sequences generated by sequential input to nonlinear sensory input systems.

The discussion is in four main parts. The first part motivates inference rather than explicit design in the implementation of NLU systems, the second outlines the proposed implementation architecture and known problems with it, the third describes Freeman's work on the dynamics of biological input systems and the relevance of his results to problems with the proposed architecture, and the fourth discusses the adequacy of the architecture in terms of generative linguistic theory.