A hybrid approach to natural language parsing

Abstract

In this paper we suggest an approach to combine principles and techniques from traditional and connectionist paradigms in order to overcome the specific problems of traditional and connectionist parsers and to simultaneously preserve their advantages. The described approach - which is deeply hybrid - is based on the idea of a dynamic generation of a neural network according to the rules of a context-free grammar and driven by the input. The neural network represents the derived parse-tree or chart of the input structure, and is involved actively in the parsing process by activation flow. The parsing is truly incremental, hypothesizes expected structures, and gives estimates for the 'probailities' of the expected structures. The hybrid parsing method has been used successfully in three different natural language processing systems (PAPADEUS, INKAS, and INKOPA).