Learning to Merge Word Senses

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 1005–1014, Prague, June 2007. c⃝2007 Association for Computational Linguistics Learning to Merge Word Senses Rion Snow Sushant Prakash Computer Science Department Stanford University Stanford, CA 94305 USA {rion,sprakash}@cs.stanford.edu Daniel Jurafsky Linguistics Department Stanford University Stanford, CA 94305 USA jurafsky@stanford.edu Andrew Y. Ng Computer Science Department Stanford University Stanford, CA 94305 USA ang@cs.stanford.edu Abstract It has been widely observed that different NLP appli- cations require different sense granularities in order to best exploit word sense distinctions, and that for many applications WordNet senses are too fine-grained. In contrast to previously proposed automatic methods for sense clustering, we formulate sense merging as a su- pervised learning problem, exploiting human-labeled sense clusterings as training data. We train a discrimi- native classifier over a wide variety of features derived from WordNet structure, corpus-based evidence, and evidence from other lexical resources. Our learned similarity measure outperforms previously proposed automatic methods for sense clustering on the task of predicting human sense merging judgments, yielding an absolute F-score improvement of 4.1% on nouns, 13.6% on verbs, and 4.0% on adjectives. Finally, we propose a model for clustering sense taxonomies us- ing the outputs of our classifier, and we make avail- able several automatically sense-clustered WordNets of various sense granularities. 1 Introduction Defining a discrete inventory of senses for a word is extremely difficult (Kilgarriff, 1997; Hanks, 2000; Palmer et al., 2005). Perhaps the greatest obstacle is the dynamic nature of sense definition: the correct granularity for word senses depends on the appli- cation. For language learners, a fine-grained set of word senses may help in learning subtle distinctions, while coarsely-defined senses are probably more useful in NLP tasks like information retrieval (Gon- zalo et al., 1998), query expansion (Moldovan and Mihalcea, 2000), and WSD (Resnik and Yarowsky, 1999; Palmer et al., 2005). Lexical resources such as WordNet (Fellbaum, 1998) use extremely fine-grained notions of word sense, which carefully capture even minor distinc- tions between different possible word senses (e.g., the 8 noun senses of bass shown in Figure 1). Pro- ducing sense-clustered inventories of arbitrary sense granularity is thus crucial for tasks which depend on lexical resources like WordNet, and is also impor- tant for the task of automatically constructing new WordNet-like taxonomies. A solution to this prob- lem must also deal with the constraints of the Word- Net taxonomy itself; for example when clustering two senses, we need to consider the transitive effects of merging synsets. The state of the art in sense clustering is insuffi- cient to meet these needs. Current sense clustering algorithms are generally unsupervised, each relying on a different set of useful features or hand-built rules. But hand-written rules have little flexibility to produce clusterings of different granularities, and previously proposed methods offer little in the di- rection of intelligently combining and weighting the many proposed features. In response to these challenges, we propose a new algorithm for clustering large-scale sense hier- archies …

Content truncated to 500 words