Using Cycles and Quasi-Cycles to Disambiguate Dictionary Glosses

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 594–602, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Using Cycles and Quasi-Cycles to Disambiguate Dictionary Glosses Roberto Navigli Dipartimento di Informatica Sapienza - Universit`a di Roma Via Salaria, 113 - 00198 Roma Italy navigli@di.uniroma1.it Abstract We present a novel graph-based algo- rithm for the automated disambiguation of glosses in lexical knowledge resources. A dictionary graph is built starting from senses (vertices) and explicit or implicit relations in the dictionary (edges). The approach is based on the identification of edge sequences which constitute cycles in the dictionary graph (possibly with one edge reversed) and relate a source to a target word sense. Experiments are per- formed on the disambiguation of ambigu- ous words in the glosses of WordNet and two machine-readable dictionaries. 1 Introduction In the last two decades, we have witnessed an increasing availability of wide-coverage lexical knowledge resources in electronic format, most notably thesauri (such as Roget’s Thesaurus (Ro- get, 1911), the Macquarie Thesaurus (Bernard, 1986), etc.), machine-readable dictionaries (e.g., the Longman Dictionary of Contemporary En- glish (Proctor, 1978)), computational lexicons (e.g. WordNet (Fellbaum, 1998)), etc. The information contained in such resources comprises (depending on their kind) sense inven- tories, paradigmatic relations (e.g. flesh3 n is a kind of plant tissue1 n),1 text definitions (e.g. flesh3 n is defined as “a soft moist part of a fruit”), usage ex- amples, and so on. Unfortunately, not all the semantics are made explicit within lexical resources. Even Word- Net, the most widespread computational lexicon of English, provides explanatory information in the form of textual glosses, i.e. strings of text 1We denote as wi p the ith sense in a reference dictionary of a word w with part of speech p. which explain the meaning of concepts in terms of possibly ambiguous words. Moreover, while computational lexicons like WordNet contain semantically explicit informa- tion such as, among others, hypernymy and meronymy relations, most thesauri, glossaries, and machine-readable dictionaries are often just elec- tronic transcriptions of their paper counterparts. As a result, for each entry (e.g. a word sense or thesaurus entry) they mostly provide implicit in- formation in the form of free text. The production of semantically richer lexical resources can help alleviate the knowledge ac- quisition bottleneck and potentially enable ad- vanced Natural Language Processing applications (Cuadros and Rigau, 2006). However, in order to reduce the high cost of manual annotation (Ed- monds, 2000), and to avoid the repetition of this effort for each knowledge resource, this task must be supported by wide-coverage automated tech- niques which do not rely on the specific resource at hand. In this paper, we aim to make explicit large quantities of semantic information implic- itly contained in the glosses of existing wide- coverage lexical knowledge resources (specifi- cally, machine-readable dictionaries and computa- tional lexicons). To this end, we present a method for Gloss Word Sense Disambiguation (WSD), called the Cycles and Quasi-Cycles (CQC) algo- rithm. The …

Content truncated to 500 words