Tor, TorMd: Distributional Profiles of Concepts for Unsupervised Word Sense Disambiguation

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 326–333, Prague, June 2007. c⃝2007 Association for Computational Linguistics TOR, TORMD: Distributional Profiles of Concepts for Unsupervised Word Sense Disambiguation Saif Mohammad Dept. of Computer Science University of Toronto Toronto, ON M5S 3G4 Canada smm@cs.toronto.edu Graeme Hirst Dept. of Computer Science University of Toronto Toronto, ON M5S 3G4 Canada gh@cs.toronto.edu Philip Resnik Dept. of Linguistics and UMIACS University of Maryland College Park, MD 20742 USA resnik@umiacs.umd.edu Abstract Words in the context of a target word have long been used as features by su- pervised word-sense classifiers. Moham- mad and Hirst (2006a) proposed a way to determine the strength of association be- tween a sense or concept and co-occurring words—the distributional profile of a con- cept (DPC)—without the use of manually annotated data. We implemented an unsu- pervised na¨ıve Bayes word sense classifier using these DPCs that was best or within one percentage point of the best unsuper- vised systems in the Multilingual Chinese– English Lexical Sample Task (task #5) and the English Lexical Sample Task (task #17). We also created a simple PMI-based classi- fier to attempt the English Lexical Substi- tution Task (task #10); however, its perfor- mance was poor. 1 Introduction Determining the intended sense of a word is poten- tially useful in many natural language tasks includ- ing machine translation and information retrieval. The best approaches for word sense disambiguation are supervised and they use words that co-occur with the target as features. These systems rely on sense- annotated data to identify words that are indicative of the use of the target in each of its senses. However, only limited amounts of sense- annotated data exist and it is expensive to create. In our previous work (Mohammad and Hirst, 2006a), we proposed an unsupervised approach to determine the strength of association between a sense or con- cept and its co-occurring words—the distributional profile of a concept (DPC)—relying simply on raw text and a published thesaurus. The categories in a published thesaurus were used as coarse senses or concepts (Yarowsky, 1992). We now show how dis- tributional profiles of concepts can be used to cre- ate an unsupervised na¨ıve Bayes word-sense classi- fier. We also implemented a simple classifier that relies on the pointwise mutual information (PMI) between the senses of the target and co-occurring words. These DPC-based classifiers participated in three SemEval 2007 tasks: the English Lexical Sam- ple Task (task #17), the English Lexical Substitu- tion Task (task #10), and the Multilingual Chinese– English Lexical Sample Task (task #5). The English Lexical Sample Task (Pradhan et al., 2007) is a traditional word sense disambiguation task wherein the intended (WordNet) sense of a tar- get word is to be determined from its context. We manually mapped the WordNet senses to the cate- gories in a thesaurus and the DPC-based na¨ıve Bayes classifier was used to identify the intended sense (category) of the target words. The object of the Lexical Substitution Task (Mc- Carthy and Navigli, 2007) …

Content truncated to 500 words