Word Sense Discrimination by Clustering Contexts in Vector and Similarity Spaces

Word Sense Discrimination by Clustering Contexts in Vector and Similarity Spaces Amruta Purandare and Ted Pedersen Department of Computer Science University of Minnesota Duluth, MN 55812 USA {pura0010,tpederse}@d.umn.edu http://senseclusters.sourceforge.net Abstract This paper systematically compares unsuper- vised word sense discrimination techniques that cluster instances of a target word that oc- cur in raw text using both vector and similarity spaces. The context of each instance is repre- sented as a vector in a high dimensional fea- ture space. Discrimination is achieved by clus- tering these context vectors directly in vector space and also by finding pairwise similarities among the vectors and then clustering in sim- ilarity space. We employ two different repre- sentations of the context in which a target word occurs. First order context vectors represent the context of each instance of a target word as a vector of features that occur in that con- text. Second order context vectors are an indi- rect representation of the context based on the average of vectors that represent the words that occur in the context. We evaluate the discrim- inated clusters by carrying out experiments us- ing sense–tagged instances of 24 SENSEVAL- 2 words and the well known Line, Hard and Serve sense–tagged corpora. 1 Introduction Most words in natural language have multiple possible meanings that can only be determined by considering the context in which they occur. Given a target word used in a number of different contexts, word sense dis- crimination is the process of grouping these instances of the target word together by determining which contexts are the most similar to each other. This is motivated by (Miller and Charles, 1991), who hypothesize that words with similar meanings are often used in similar contexts. Hence, word sense discrimination reduces to the problem of finding classes of similar contexts such that each class represents a single word sense. Put another way, contexts that are grouped together in the same class represent a particular word sense. While there has been some previous work in sense dis- crimination (e.g., (Sch¨utze, 1992), (Pedersen and Bruce, 1997), (Pedersen and Bruce, 1998), (Sch¨utze, 1998), (Fukumoto and Suzuki, 1999)), by comparison it is much less than that devoted to word sense disambiguation, which is the process of assigning a meaning to a word from a predefined set of possibilities. However, solutions to disambiguation usually require the availability of an external knowledge source or manually created sense– tagged training data. As such these are knowledge inten- sive methods that are difficult to adapt to new domains. By contrast, word sense discrimination is an unsuper- vised clustering problem. This is an attractive methodol- ogy because it is a knowledge lean approach based on ev- idence found in simple raw text. Manually sense tagged text is not required, nor are specific knowledge rich re- sources like dictionaries or ontologies. Instances are clus- tered based on their mutual contextual similarities which can be completely computed from the text itself. This paper presents a systematic comparison of dis- crimination techniques suggested …

Content truncated to 500 words