LIMSI : Cross-lingual Word Sense Disambiguation using Translation Sense Clustering

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 178–182, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics LIMSI : Cross-lingual Word Sense Disambiguation using Translation Sense Clustering Marianna Apidianaki LIMSI-CNRS Rue John Von Neumann 91403 Orsay Cedex, France marianna@limsi.fr Abstract We describe the LIMSI system for the SemEval-2013 Cross-lingual Word Sense Dis- ambiguation (CLWSD) task. Word senses are represented by means of translation clusters in different languages built by a cross-lingual Word Sense Induction (WSI) method. Our CLWSD classifier exploits the WSI output for selecting appropriate translations for target words in context. We present the design of the system and the obtained results. 1 Introduction This paper describes the LIMSI system that partici- pated in the Cross-Lingual Word Sense Disambigua- tion (CLWSD) task of SemEval-2013. The goal of CLWSD is to predict semantically correct transla- tions for ambiguous words in context (Resnik and Yarowsky, 2000; Carpuat and Wu, 2007; Apidi- anaki, 2009). The CLWSD task of the SemEval-2013 evaluation campaign is a lexical sample task for En- glish nouns and is divided into two subtasks: the best subtask where systems are asked to provide a unique good translation for words in context; the out-of-five (oof) subtask where systems can propose up to five semantically related translations for each target word instance (Lefever and Hoste, 2013). The CLWSD lexical sample contains 20 nouns and the test set is composed of 50 instances per noun. Sys- tem performance is evaluated by comparing the sys- tem output to a set of gold standard annotations in five languages: French, Spanish, Italian, Dutch and German. Participating systems have to provide con- textually appropriate translations for target words in context in each or a subset of the target languages. We apply the CLWSD method proposed by Apid- ianaki (2009) to three bilingual tasks: English- Spanish, English-French and English-Italian. The method exploits the translation clusters generated in the three target languages by a cross-lingual Word Sense Induction (WSI) method. The WSI method clusters the translations of target words in a parallel corpus using source language context vectors. The same vectors are exploited during disambiguation in order to select the most appropriate translations for new instances of the target words in context. 2 System Description 2.1 Translation clustering Contrary to monolingual WSI methods which group the instances of the words into clusters describ- ing their senses, the cross-lingual WSI method used here clusters the translations of words in a paral- lel corpus. The corpus used for French consists of the English-French parts of Europarl (version 7) (Koehn, 2005) and of the JRC-Acquis corpus (Stein- berger et al., 2006), joined together. For English- Spanish and English-Italian we only use the corre- sponding parts of Europarl. The corpora are first tokenized and lowercased using the Moses scripts, then lemmatized and tagged by part-of-speech (PoS) using the TreeTagger (Schmid, 1994). Words in the corpus are replaced by a lemma and PoS tag pair be- fore word alignment, …

Content truncated to 500 words