Regular polysemy: A distributional model

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 151–160, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Regular polysemy: A distributional model Gemma Boleda Dept. of Linguistics University of Texas at Austin gemma.boleda@upf.edu Sebastian Pad´o ICL University of Heidelberg pado@cl.uni-heidelberg.de Jason Utt IMS University of Stuttgart uttjn@ims.uni-stuttgart.de Abstract Many types of polysemy are not word specific, but are instances of general sense alternations such as ANIMAL-FOOD. Despite their perva- siveness, regular alternations have been mostly ignored in empirical computational semantics. This paper presents (a) a general framework which grounds sense alternations in corpus data, generalizes them above individual words, and allows the prediction of alternations for new words; and (b) a concrete unsupervised implementation of the framework, the Cen- troid Attribute Model. We evaluate this model against a set of 2,400 ambiguous words and demonstrate that it outperforms two baselines. 1 Introduction One of the biggest challenges in computational se- mantics is the fact that many words are polysemous. For instance, lamb can refer to an animal (as in The lamb squeezed through the gap) or to a food item (as in Sue had lamb for lunch). Polysemy is pervasive in human language and is a problem in almost all applications of NLP, ranging from Machine Trans- lation (as word senses can translate differently) to Textual Entailment (as most lexical entailments are sense-specific). The field has thus devoted a large amount of effort to the representation and modeling of word senses. The arguably most prominent effort is Word Sense Disambiguation, WSD (Navigli, 2009), an in-vitro task whose goal is to identify which, of a set of pre- defined senses, is the one used in a given context. In work on WSD and other tasks related to pol- ysemy, such as word sense induction, sense alter- nations are treated as word-specific. As a result, a model for the meaning of lamb that accounts for the relation between the animal and food senses cannot predict that the same relation holds between instances of chicken or salmon in the same type of contexts. A large number of studies in linguistics and cog- nitive science show evidence that there are regulari- ties in the way words vary in their meaning (Apres- jan, 1974; Lakoff and Johnson, 1980; Copestake and Briscoe, 1995; Pustejovsky, 1995; Gentner et al., 2001; Murphy, 2002), due to general analogical processes such as regular polysemy, metonymy and metaphor. Most work in theoretical linguistics has focused on regular, systematic, or logical polysemy, which accounts for alternations like ANIMAL-FOOD. Sense alternations also arise from metaphorical use of words, as dark in dark glass-dark mood, and also from metonymy when, for instance, using the name of a place for a representative (as in Germany signed the treatise). Disregarding this evidence is empiri- cally inadequate and leads to the well-known lexical bottleneck of current word sense models, which have serious problems in achieving high coverage (Navigli, 2009). We believe that empirical computational semantics could profit from a model of polysemy1 which …

Content truncated to 500 words