Syntax-Semantics Interaction in Sentence Understanding

Thumbnail Image
Mahesh, Kavi
Associated Organizations
Organizational Unit
Supplementary to
Natural language is the primary mode of human communication. Developing a complete and well-specified computational model of language understanding is a difficult problem. Understanding a natural language sentence requires the application of many types of knowledge, such as syntactic, semantic, and conceptual knowledge, to resolve the many types of ambiguities that abound in natural language. Most unresolved issues in both psychological and computational modeling of sentence understanding are concerned with the questions of when should each of the various types of knowledge be applied in processing a sentence and how should the different types of knowledge be integrated to select unique interpretations of sentences. In this work, we have developed a model of sentence understanding called COMPERE (Cognitive Model of Parsing and Error Recovery). Our model was built on the hypothesis that a sentence processor has an architecture with separate representations of the different types of knowledge but a single unified process that integrates the different types of knowledge. We have shown that such an architecture addresses the modularity debate by demonstrating how the same sentence processor can produce seemingly modular behaviors in some situations and interactive behaviors in other situations. We have also shown how the unified arbitrating process can not only resolve both syntactic and semantic, lexical and structural, ambiguities, but can also recover from its errors in both syntactic and semantic ambiguity resolution. The unified process can also explain the temporal dependencies in syntax-semantics interactions. It shows how certain decisions are made early and others delayed until further information becomes available. We have developed a parsing algorithm called Head-Signaled Left-Corner parsing to identify the time course of points in the sentence where decisions are to be made. This algorithm decides when to make a commitment and when to delay a syntactic attachment. We have also developed a simple arbitration algorithm for combining information coming from multiple knowledge sources and for resolving any conflicts between them. In addition we have developed a uniform representation of syntactic and semantic interpretations using what are called intermediate roles. These intermediate roles not only aid the dynamic integration of knowledge types by the unified arbitrator, they also provide a declarative record of the intermediate decisions made in syntax-semantics interactions to enable the processor to recover from its errors through repair rather than complete reprocessing. We present a theoretical framework for formal analyses of the performance of sentence processors in various situations. These analyses indicate that the HSLC parsing algorithm, along with incremental interactions between syntax and semantics controlled by the unified arbitrator, reduces the amount of local ambiguity and working memory requirements in processing a sentence. We also present certain psychological predictions made by the COMPERE model.
Date Issued
1297180 bytes
Resource Type
Resource Subtype
Technical Report
Rights Statement
Rights URI