An Integrated Approach to Measuring Semantic Similarity between Words Using Information Available on the Web

Proceedings of NAACL HLT 2007, pages 340–347, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics An Integrated Approach to Measuring Semantic Similarity between Words Using Information available on the Web Danushka Bollegala The University of Tokyo 7-3-1, Hongo, Tokyo, 113-8656, Japan danushka@mi.ci.i.u- tokyo.ac.jp Yutaka Matsuo National Institute of Advanced Industrial Science and Technology 1-18-13, Sotokanda, Tokyo, 101-0021, Japan y.matsuo@aist.go.jp Mitsuru Ishizuka The University of Tokyo 7-3-1, Hongo, Tokyo, 113-8656, Japan ishizuka@i.u- tokyo.ac.jp Abstract Measuring semantic similarity between words is vital for various applications in natural language processing, such as language modeling, information retrieval, and document clustering. We propose a method that utilizes the information avail- able on the Web to measure semantic sim- ilarity between a pair of words or entities. We integrate page counts for each word in the pair and lexico-syntactic patterns that occur among the top ranking snippets for the AND query using support vector ma- chines. Experimental results on Miller- Charles’ benchmark data set show that the proposed measure outperforms all the ex- isting web based semantic similarity mea- sures by a wide margin, achieving a cor- relation coefficient of 0.834. Moreover, the proposed semantic similarity measure significantly improves the accuracy (F- measure of 0.78) in a named entity cluster- ing task, proving the capability of the pro- posed measure to capture semantic simi- larity using web content. 1 Introduction The study of semantic similarity between words has been an integral part of natural language processing and information retrieval for many years. Semantic similarity measures are vital for various applications in natural language processing such as word sense disambiguation (Resnik, 1999), language model- ing (Rosenfield, 1996), synonym extraction (Lin, 1998a) and automatic thesaurus extraction (Curran, 2002). Pre-compiled taxonomies such as WordNet 1 and text corpora have been used in previous work on se- mantic similarity (Lin, 1998a; Resnik, 1995; Jiang and Conrath, 1998; Lin, 1998b). However, seman- tic similarity between words change over time as new senses and associations of words are constantly created. One major issue behind taxonomies and corpora oriented approaches is that they might not necessarily capture similarity between proper names such as named entities (e.g., personal names, loca- tion names, product names) and the new uses of ex- isting words. For example, apple is frequently asso- ciated with computers on the Web but this sense of apple is not listed in the WordNet. Maintaining an up-to-date taxonomy of all the new words and new usages of existing words is costly if not impossible. The Web can be regarded as a large-scale, dy- namic corpus of text. Regarding the Web as a live corpus has become an active research topic recently. Simple, unsupervised models have shown to per- form better when n-gram counts are obtained from the Web rather than from a large corpus (Keller and Lapata, 2003; Lapata and Keller, 2005). Resnik and Smith (2003) extract bilingual sentences from the Web to create parallel corpora for machine trans- lation. Turney (2001) defines a point wise mutual information (PMI-IR) measure using the …

Content truncated to 500 words