Unsupervised Word Sense Induction from Multiple Semantic Spaces with Locality Sensitive Hashing

International Conference RANLP 2009 - Borovets, Bulgaria, pages 287–291 Unsupervised Word Sense Induction from Multiple Semantic Spaces with Locality Sensitive Hashing Claire Mouton - CEA LIST - Exalead CEA LIST - Claire.Mouton@cea.fr Exalead - 10 place de la Madeleine - 75008 Paris Claire.Mouton@exalead.com Guillaume Pitel - Epiphyte 10 rue de Vouill´e - 75015 Paris Guillaume.Pitel@epiphyte.eu Ga¨el de Chalendar - CEA LIST 18, route du Panorama BP6, FONTENAY AUX ROSES F-92265 France Gael.de-Chalendar@cea.fr Anne Vilnat - LIMSI/CNRS bˆat. 508 - Universit´e Paris XI BP 133 - 91403 Orsay Cedex France Anne.Vilnat@limsi.fr Abstract Word Sense Disambiguation is the task dedi- cated to the problem of finding out the sense of a word in context, from all of its many pos- sible senses. Solving this problem requires to know the set of possible senses for a given word, which can be acquired from human knowledge, or from automatic discovery, called Word Sense Induction. In this article, we adapt two exist- ing meta-methods of Word Sense Induction for the automatic construction of a disambiguation lexicon. Our adaptation is based on multiple se- mantic spaces (also called Word Space Models) produced from a syntactic analysis of a very large number of web pages. These adaptations and the results presented in this article differ from the original methods in that they use a combination of several high dimensional spaces instead of one single representation. Each of these competing semantic spaces takes part in a clustering phase in which they vote on sense induction. Keywords semantic space, word space model, dimensionality reduction, lo- cality sensitive hashing, word sense induction, words clustering, multi-represented data 1 Introduction A single word may potentially convey many different senses but despite this potential for ambiguity, misin- terpretations occur relatively rarely in actual human linguistic interactions, including written texts. As a consequence, it is possible to draw the hypothesis that a wordsense should be inferable from the context it appears in (letting aside the question of the context size). Word Sense Disambiguation is the task dedi- cated to this problem of finding out the sense of a word in context. For a more detailed study, [1] and [12] have covered an exhausting overview of the task. However, solving this problem requires first to know the set of possible senses for a given word. The mere possibility of doing this exhaustively is a much debated issue, because even among humans, it is very difficult to agree on a set of senses for any given word. How- ever if we consider it possible to obtain a list of the senses, even if it is incomplete, then it can be acquired from human knowledge (as one can find in any good dictionary), or from automatic discovery, a task which is generally called Word Sense Induction. This task (and consequently Word Sense Disam- biguation itself) is often tackled using methods that are able to group similar words together, in order to perform a clustering over the neighbors of the target, and thus to discover one ’sense’ for …

Content truncated to 500 words