Walk-based Computation of Contextual Word Similarity

Proceedings of COLING 2012: Technical Papers, pages 1081–1096, COLING 2012, Mumbai, December 2012. Walk-based Computation of Contextual Word Similarity Kazuo Hara1 Ikumi Suzuki1 Masashi Shimbo2 Yuji Matsumoto2 (1) National Institute of Genetics, Mishima, Shizuoka, Japan (2) Nara Institute of Science and Technology, Ikoma, Nara, Japan kazuo.hara@gmail.com, suzuki.ikumi@gmail.com, shimbo@is.naist.jp, matsu@is.naist.jp ABSTRACT We propose a new measure of semantic similarity between words in context, which exploits the syntactic/semantic structure of the context surrounding each target word. For a given pair of target words and their sentential contexts, labeled directed graphs are made from the output of a semantic parser on these sentences. Nodes in these graphs represent words in the sentences, and labeled edges represent syntactic/semantic relations between them. The similarity between the target words is then computed as the sum of the similarity of walks starting from the target words (nodes) in the two graphs. The proposed measure is tested on word sense disambiguation and paraphrase ranking tasks, and the results are promising: The proposed measure outperforms existing methods which completely ignore or do not fully exploit syntactic/semantic structural co-occurrences between a target word and its neighbors. KEYWORDS: contextual word similarity, word sense disambiguation, paraphrase, kernel method. 1081 1 Introduction Word ambiguity poses a great difficulty in natural language processing (Deerwester et al., 1990; Berger and Lafferty, 1999; Navigli, 2009). In document classification, for example, polysemous words may provide spurious evidence for misclassifying documents of different categories into a single category. Synonyms may also cause problems; without the knowledge of two words being synonyms, classifiers might be unable to detect similarity of documents in which they appear. These problems can be alleviated with the help of contextual similarity (Jurafsky and Martin, 2008, Section 20.7). Polysemous words have identical surface forms, but the contexts in which they appear are often distinct; and even though synonyms have different surface forms, they are expected to appear in similar context. There has been a volume of work investigating semantic similarity of words in context. However, syntactic or semantic structure in the context has not been fully explored. The methods representing context as a bag of words or a bag of n-grams (Schütze, 1998; Reisinger and Mooney, 2010; Erk and Padó, 2010; Dinu and Lapata, 2010; Giuliano et al., 2009) ignore the underlying syntactic/semantic structure. Recent studies on compositional semantics of words (Mitchell and Lapata, 2008; Erk and Padó, 2008; Thater et al., 2010) take advantage of structure in context, but they only use information of words that directly stand in a predicate-argument relation with the target word. Consider the following sentences (i) and (ii), shown with the word dependency relations. (i) The branch is referred to as the face of the bank. (ii) These plants shielded the face of the bank. In sentence (i), the noun (homonym) bank means “a financial institution” whereas the one in sentence (ii) means “the slope of land adjoining a river.” If these sentences are treated as bags of words, we see that the words unique to (i) and (ii) …

Content truncated to 500 words