HERMIT: Flexible Clustering for the SemEval-2 WSI Task

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 359–362, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics HERMIT: Flexible Clustering for the SemEval-2 WSI Task David Jurgens University of California, Los Angeles Los Angeles, California, USA jurgens@cs.ucla.edu Keith Stevens University of California, Los Angeles Los Angeles, California, USA kstevens@cs.ucla.edu Abstract A single word may have multiple un- specified meanings in a corpus. Word sense induction aims to discover these dif- ferent meanings through word use, and knowledge-lean algorithms attempt this without using external lexical resources. We propose a new method for identify- ing the different senses that uses a flexi- ble clustering strategy to automatically de- termine the number of senses, rather than predefining it. We demonstrate the effec- tiveness using the SemEval-2 WSI task, achieving competitive scores on both the V-Measure and Recall metrics, depending on the parameter configuration. 1 Introduction The Word Sense Induction task of SemEval 2010 compares several sense induction and discrimina- tion systems that are trained over a common cor- pus. Systems are provided with an unlabeled train- ing corpus consisting of 879,807 contexts for 100 polysemous words, with 50 nouns and 50 verbs. Each context consists of several sentences that use a single sense of a target word, where at least one sentence contains the word. Systems must use the training corpus to induce sense representations for the many word senses and then use those represen- tations to produce sense labels for the same 100 words in unseen contexts from a testing corpus. We perform this task by utilizing a distribu- tional word space formed using dimensionality reduction and a hybrid clustering method. Our model is highly scalable; the dimensionality of the word space is reduced immediately through a pro- cess based on random projections. In addition, an online part of our clustering algorithm maintains only a centroid that describes an induced word sense, instead of all observed contexts, which lets the model scale to much larger corpora than those used in the SemEval-2 WSI task. 2 The Word Sense Induction Model We perform word sense induction by modeling individual contexts in a high dimensional word space. Word senses are induced by finding con- texts which are similar and therefore likely to use the same sense of the target word. We use a hybrid clustering method to group similar contexts. 2.1 Modeling Context For a word, each of its contexts are represented by the words with which it co-occurs. We approx- imate this high dimensional co-occurrence space with the Random Indexing (RI) word space model (Kanerva et al., 2000). RI represents the occur- rence of a word with an index vector, rather than a set of dimensions. An index vector is a fixed, sparse vector that is orthogonal to all other words’ index vectors with a high probability; the total number of dimensions in the model is fixed at a small value, e.g. 5,000. Orthogonality is obtained by setting a small percentage of the vector’s values to ±1 …

Content truncated to 500 words