SGER: Sense in Context: Construction of a Selectional Context Dictionary for English Clauses

Brandeis University, Waltham MA

Investigators

Abstract

ABSTRACT This research explores a novel framework for constructing a context dictionary for natural language; that is, a computational lexical database of selectional contexts for argument-taking predicators in the language. The goal of this project is to semi-automatically identify, from very large corpora of the language, the norms of word usage, and then to encode these norms as selectional contexts using a restricted system of semantic types and syntactic identifiers. A novel methodology, called Corpus Pattern Analysis (CPA), which is quite different from established methods in LKB construction, will be used. Rather than building MRD-seeded lexicons, hand-crafted lexicons (WordNet, EuroWordNet), and corpus-informed lexicons (FrameNet), this work will employ a methodology for building a "corpus-driven" lexicon, constrained by specific linguistic modeling principles. Such a "selectional context dictionary" overcomes many of the shortcomings of previous LKB approaches by capturing a richer level of selectional context for predicates in the language. This is possible with the use of shallow parsing of predicate arguments, and the assignment of types from a shallow semantic type system. The technique will be evaluated the selectional discriminatory capabilities of the results, according to established methods; namely, precision and recall measurements against an annotated test set of unseen text.