Developing and applying an integrated semantic framework for natural language understanding

Abstract

The standard approach in Natural Language Processing for semantic analysis (Word-Sense Disambiguation, Named-Entity Recognition and other related tasks) is to match tokens from shallow parsed text (tokenized, POS tagged, shallow trunking, et cetera) to a sense repository and then rank them to find the best candidates. This practice has yet to exploit the extra information that is available in structural semantics, which can be accessed using deep grammars. This dissertation proposes the Integrated Semantic Framework, a novel method to improve computational semantic analysis by providing both structural semantics from construction grammars and lexical semantics from ontologies in a single representation. The method was implemented as a software package that produces computational semantic analysis and its performance was compared to human annotators and some others semantic analysis systems on a short story. Currently, the implemented system only provides analyses for standard English texts. However the design is extensible to other languages and is already being developed for the Japanese language. Finally, although the implemented system is still a prototype (most rules are generated automatically, the structure matching and transforming features are still at a basic level, and a few other tasks remain on the to-improve list), the results prove that such a system can be built and can produce positive results. This research demonstrated that it is possible to provide a more natural and sophisticated computational semantic analysis. It aims to motivate linguists to join the development of fundamental semantic theories in the field of Natural Language Processing; to interpret and provide better semantics that exist in natural languages.