The Acquisition of Lexical Knowledge from Combined Machine-Readable Dictionary Sources
چکیده
This paper is concerned with the question of how to extract lexical knowledge from MachineReadable Dictionaries (MRDs) within a lexical database which integrates a lexicon development environment. Our long te rm objective is the creation of a large lexical knowledge base using semiautomatic techniques to recover syntactic and semantic information from MRDs. In doing so, one finds that reliance on a single MRD source induces inadequacies which could be efficiently redressed through access to combined MRD sources. In the general case, the integration of information from distinct MRDs remains a problem hard, perhaps impossible, to solve without the aid of a complete, linguistically motivated database which provides a reference point for comparison. Nevertheless, advances can be made by a t tempt ing to correlate dictionaries which are not too dissimilar. In keeping with these observations, we describe a software package for correlating MRDs based on sense merging techniques and show how such a tool can be employed in augmenting a lexical knowledge base built f rom a conventional MRD with thesaurus information. 1 I n t r o d u c t i o n Over the last few years, the utilization of machine readable dictionaries (MRDs) in compiling lexical components for Natural Language Processing (NLP) systems has awakened the interest of an increasing number of researchers. This trend has largely arisen from the recognition tha t access to a large scale Lexical Knowledge Base (LKB) is a s i n e qua n o n for real world applications of NLP systems. Computer-aided acquisition of lexical knowledge from MRDs can be made to satisfy this prerequisite in a manner which is both timeand costeffective. However, to maximize the utility of MRDs in *The research reported in this paper was carried out at the Computer Laboratory (University of Cambridge) as part of an ongoing study on the acquisition of lexical knowledge from machine-readable dictionaries within the ACQUILEX project, ESPRIT BRA 3030. NLP applications, it is necessary to overcome inconsistencies, omissions and the occasional errors which are commonly found in MRDs (Atkins el al., 1986; Atkins, 1989; Akkerman, 1989; Boguraev & Briscoe, 1989). This goal can be par t ly achieved by developing tools which make it possible to correct errors and inconsistencies contained in the information structures automatically derived from the source MRD before these are stored in the LKB or target lexicon (Carroll & Grover, 1989). This technique is nevertheless of little avail in redressing inadequacies which arise from lack of information. In this case, manual supply of the missing information would be too timeand labour-intensive to be desirable. Moreover, the information which is missing can be usually obtained from other MRD sources. Consider, for example, a situation in which we wanted to augment lexical representations available from a conventional dictionary with thesaurus information. There can be little doubt tha t the integration of information f rom distinct MRD sources would be far more convenient and appropriate than reliance on manual encoding. In the general case, the integration of information from distinct MRD sources for use within a lexicon development environment is probably going to remain an unsolved problem for quite some time. This is simply because dictionaries seldom describe the same word using the same sense distinctions. Consequently, the integration of information from distinct MRD sources through simple word-sense matches is likely to fail in a significant number of instances (Calzolari & Picchi, 1986; Atkins 1987; Klavans 1988; Boguraev & Pustejovsky 1990). Indeed, Atkins & Levin (1990) have suggested that the task of mapping MRDs onto each other is so complex that the creation of a complete ' ideal ' database which provides a reference point for the MRD sources to be integrated may well be considered as an essential prerequisite. However, when dealing with MRD sources which use entry definitions which are not too dissimilar, a correlation technique based on word sense merging can be made to yield useful results, given the appropriate tools Although sense matching across dictionaries in this case too is prone to errors, there are several reasons why the effort is worthwhile. First, the number of correct sens~ matches across MRD sources in this case is guarantee~ to be high. Second, there are many instances in which at incorrect sense-to-sense match does not affect the fins