Extraction of Thematic Roles from Dictionary Definitions

Language, Information and Computation(PACLIC 11),1996, 137-146 Extraction of Thematic Roles from Dictionary Definitions Dr. Michael L. Mc Hale RL/C3CA Rome Laboratory 525 Brooks Road Rome, New York 13441-4505 mchale@ai.rl.af.mil Prof. Sung H. Myaeng Department of Computer Science Chungnam National University 220 Kung-dong, Yoosung-ku Taejon, 305-764 Korea shmyaeng@cs.chungnam.ac.kr Abstract Our research goal has been the development of a domain independent natural language processing (NLP) system suitable for information retrieval. As part of that research, we have investigated ways to automatically extend the semantics of a lexicon derived from machine-readable lexical sources. This paper details the extraction of thematic roles derived from lexical patterns in a machine-readable dictionary. Introduction With information retrieval as a goal, we are very sensitive to the issues of generalization, speed and scalability. Any NLP system that is used for information retrieval must be capable of handling large amounts of general text in a timely manner. Each of the components of such a system, from morphology through semantics, must have similar capabilities. Thematic roles, which provide a very basic "who does what to whom" type of semantics, meet these criteria because they are very general, simple and useable by a wide variety of natural language processing systems. The roles are generally contained in frames that contain the type of each argument for each verb. For example eat would have a frame similar to eat[AGENT, THEME]. We have explored the development of these frames by using information found in an on-line version of Longman's Dictionary of Contemporary English (LDOCE 1987). The information in the dictionary includes: definitions; subject field codes; the box codes, that provide information on the type of arguments (ex., human or abstract); a reduced set of grammar codes, that provide information on the transitivity of a verb and the syntactic category of any extra arguments; .and other information much of which is probably extraneous to the roles. 137 We focused on the definitions because we felt that they would provide the best base for a scaleable approach. Not only do the definitions contain information germane to role extraction but they are complete in the sense that every verb has a definition. This is not the case for the other types of information available. The box codes, for instance, are generic (i.e., empty) codes around 17% of the time. Our approach to analyzing the definitions is based on lexical patterns. A lexical pattern is simply a series of consecutive words that is used in more than one definition. Some of these have the appearance of a syntactic pattern (ex., to cause to) while others more directly reflect their lexical nature (ex., longer have because). Lexical patterns are the logical place to start because they are the lowest level of analysis that seems likely to contain sufficient information for role extraction. Obviously, we could have included tagging, syntactic analysis, syntactic patterns, statistics on the box codes or some other information. By focusing on the lowest level of processing, however, we ensure that any information that is extraneous to …

Content truncated to 500 words