UMND2 : SenseClusters Applied to the Sense Induction Task of Senseval-4

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 394–397, Prague, June 2007. c⃝2007 Association for Computational Linguistics UMND2 : SenseClusters Applied to the Sense Induction Task of SENSEVAL-4 Ted Pedersen Department of Computer Science University of Minnesota Duluth, MN 55812 tpederse@d.umn.edu http://senseclusters.sourceforge.net Abstract SenseClusters is a freely–available open– source system that served as the Univer- sity of Minnesota, Duluth entry in the SENSEVAL-4 sense induction task. For this task SenseClusters was configured to con- struct representations of the instances to be clustered using the centroid of word co- occurrence vectors that replace the words in an instance. These instances are then clustered using k–means where the number of clusters is discovered automatically using the Adapted Gap Statistic. In these experi- ments SenseClusters did not use any infor- mation outside of the raw untagged text that was to be clustered, and no tuning of the sys- tem was performed using external corpora. 1 Introduction The object of the sense induction task of SENSEVAL-4 (Agirre and Soroa, 2007) was to cluster 27,132 instances of 100 different words (35 nouns and 65 verbs) into senses or classes. The task data consisted of the combination of the test and training data (minus the sense tags) from the English lexical sample task. Each instance is a context of several sentences which contains an occurrence of a given word that serves as the target of sense induction. SenseClusters is based on the presumption that words that occur in similar contexts will have similar meanings. This intuition has been presented as both the Distributional Hypothesis (Harris, 1968) and the Strong Contextual Hypothesis (Miller and Charles, 1991). SenseClusters has been in active development at the University of Minnesota, Duluth since 2002. It is an open–source project that is freely–available from sourceforge, and has been been described in detail in numerous publications (e.g., (Purandare and Ped- ersen, 2004), (Pedersen et al., 2005), (Pedersen and Kulkarni, 2007)). SenseClusters supports a variety of techniques for selecting lexical features, representing contexts to be clustered, determining the appropriate number of cluster automatically, clustering, labeling of clus- ters, and evaluating cluster quality. The configu- ration used in SENSEVAL-4 was just one possible combination of these techniques. 2 Methodology in Sense Induction Task For this task, SenseClusters represents the instances to be clustered using second order co–occurrence vectors. These are constructed by first identifying word co–occurrences, and then replacing each word in an instance to be clustered with its co-occurrence vector. Then all the vectors that make up an instance are averaged together to represent that instance. A co–occurrence matrix is constructed by identi- fying bigrams that occur in the contexts to be clus- tered two or more times and have a Pointwise Mu- tual Information (PMI) score greater than five. If the value of PMI is near 1.0, this means that the words in the bigram occur together approximately the num- ber of times expected by chance, and they are not strongly associated. If this value is greater than 1, then …

Content truncated to 500 words