Generating a Parsing Lexicon From Lexical-Conceptual Structure

摘要

This paper describes the generation of a lexicon for a principle- based parser (Minipar 5,%1) using descriptions of verbs based on Lexical-Conceptual Structure (LCS 1,21). We aim to determine how much syntactic information we can obtain from a semantic-based lexicon. More specifically, we aim to provide a general approach to projection of syntactic entries from semantic (language-independent) lexicons-and to test the effect of such lexicons on parser performance. Verbs are grouped together into classes-each denoted by an LCS representation adn the thematic grid. These are mapped systematically into syntactic categories associated with entries in the Minipar parser. The main advantage of this LCS-to-syntax projection is language potability: We currently have LCS lexicons for English, Arabic, Spanish, and Chinese; thus, our LCS-projection approach allows us to produce syntactic lexicons for parsing in each of these languages. For comparing the performance of the projection from the LCS to Minipar coes, we also generated the mappings for the codes of Longman's Directionary of Contemporary English (LDOCE 81)-the most comprehensive online dictionary for syntactic categorization. Preliminary experiments indicate that our approach yields a categorization of verbs with 58% precision and 65% recall as measured against LDOCE-with an improved precision of 74% when redundancy is removed. The next section presents a brief description of each code set we use. In Section 3, we explain how we generated Minipar codes from LCS representation. Finally, Section 4, discusses our experiments and results.