UOY: A Hypergraph Model For Word Sense Induction & Disambiguation

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 414–417, Prague, June 2007. c⃝2007 Association for Computational Linguistics UOY: A Hypergraph Model For Word Sense Induction & Disambiguation Ioannis P. Klapaftis University of York Department of Computer Science giannis@cs.york.ac.uk Suresh Manandhar University of York Department of Computer Science suresh@cs.york.ac.uk Abstract This paper is an outcome of ongoing re- search and presents an unsupervised method for automatic word sense induction (WSI) and disambiguation (WSD). The induction algorithm is based on modeling the co- occurrences of two or more words using hypergraphs. WSI takes place by detect- ing high-density components in the co- occurrence hypergraphs. WSD assigns to each induced cluster a score equal to the sum of weights of its hyperedges found in the lo- cal context of the target word. Our system participates in SemEval-2007 word sense in- duction and discrimination task. 1 Introduction The majority of both supervised and unsupervised approaches to WSD is based on the “fixed-list” of senses paradigm where the senses of a target word is a closed list of definitions coming from a stan- dard dictionary (Agirre et al., 2006). Lexicographers have long warned about the problems of such an ap- proach, since dictionaries are not suited to this task; they often contain general definitions, they suffer from the lack of explicit semantic and topical rela- tions or interconnections, and they often do not re- flect the exact content of the context, in which the target word appears (Veronis, 2004). To overcome this limitation, unsupervised WSD has moved towards inducing the senses of a target word directly from a corpus, and then disambiguat- ing each instance of it. Most of the work in WSI is based on the vector space model, where the con- text of each instance of a target word is represented as a vector of features (e.g second-order word co- occurrences) (Schutze, 1998; Purandare and Peder- sen, 2004). These vectors are clustered and the re- sulting clusters represent the induced senses. How- ever, as shown experimentally in (Veronis, 2004), vector-based techniques are unable to detect low- frequency senses of a target word. Recently, graph-based methods were employed in WSI to isolate highly infrequent senses of a target word. HyperLex (Veronis, 2004) and the adaptation of PageRank (Brin and Page, 1998) in (Agirre et al., 2006) have been shown to outperform the most fre- quent sense (MFS) baseline in terms of supervised recall, but they still fall short of supervised WSD systems. Graph-based approaches operate on a 2- dimensional space, assuming a one-to-one relation- ship between co-occurring words. However, this assumption is insufficient, taking into account the fact that two or more words are usually combined to form a relationship of concepts in the context. Additionally, graph-based approaches fail to model and exploit the existence of collocations or terms consisting of more than two words. This paper proposes a method for WSI, which is based on a hypergraph model operating on a n-dimensional space. In such a model, co- occurrences of …

Content truncated to 500 words