SenseSpotting: Never let your parallel data tie you to an old domain

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1435–1445, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics SenseSpotting: Never let your parallel data tie you to an old domain Marine Carpuat1, Hal Daum´e III2, Katharine Henry3, Ann Irvine4, Jagadeesh Jagarlamudi5, Rachel Rudinger6 1 National Research Council Canada, marine.carpuat@nrc.gc.ca 2 CLIP, University of Maryland, me@hal3.name 3 CS, University of Chicago, kehenry@uchicago.edu 4 CLSP, Johns Hopkins University, anni@jhu.edu 5 IBM T.J. Watson Research Center, jags@us.ibm.com 6 CLSP, Johns Hopkins University, rachel.rudinger@aya.yale.edu Abstract Words often gain new senses in new do- mains. Being able to automatically iden- tify, from a corpus of monolingual text, which word tokens are being used in a pre- viously unseen sense has applications to machine translation and other tasks sensi- tive to lexical semantics. We define a task, SENSESPOTTING, in which we build sys- tems to spot tokens that have new senses in new domain text. Instead of difficult and expensive annotation, we build a gold- standard by leveraging cheaply available parallel corpora, targeting our approach to the problem of domain adaptation for ma- chine translation. Our system is able to achieve F-measures of as much as 80%, when applied to word types it has never seen before. Our approach is based on a large set of novel features that capture varied aspects of how words change when used in new domains. 1 Introduction As Magnini et al. (2002) observed, the domain of the text that a word occurs in is a useful signal for performing word sense disambiguation (e.g. in a text about finance, bank is likely to refer to a finan- cial institution while in a text about geography, it is likely to refer to a river bank). However, in the classic WSD task, ambiguous word types and a set of possible senses are known in advance. In this work, we focus on the setting where we observe texts in two different domains and want to iden- tify words in the second text that have a sense that did not appear in the first text, without any lexical knowledge in the new domain. To illustrate the task, consider the French noun rapport. In the parliament domain, this means ´etat rapport r´egime Govt. geo. state report (political) regime Medical state (mind) report diet geo. state ratio (political) regime Science geo. state ratio (political) regime report diet Movies geo. state report (political) regime diet Table 1: Examples of French words and their most frequent senses (translations) in four domains. (and is translated as) “report.” However, in mov- ing to a medical or scientific domain, the word gains a new sense: “ratio”, which simply does not exist in the parliament domain. In a science do- main, the “report” sense exists, but it is dominated about 12:1 by “ratio.” In a medical domain, the “report” sense remains dominant (about 2:1), but the new “ratio” sense appears frequently. In this paper we define a new task that we call SENSESPOTTING. The goal of this …

Content truncated to 500 words