Revisiting Context-based Projection Methods for Term-Translation Spotting in Comparable Corpora

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 617–625, Beijing, August 2010 Revisiting Context-based Projection Methods for Term-Translation Spotting in Comparable Corpora Audrey Laroche OLST – D´ep. de linguistique et de traduction Universit´e de Montr´eal audrey.laroche@umontreal.ca Philippe Langlais RALI – DIRO Universit´e de Montr´eal felipe@iro.umontreal.ca Abstract Context-based projection methods for identifying the translation of terms in comparable corpora has attracted a lot of attention in the community, e.g. (Fung, 1998; Rapp, 1999). Surprisingly, none of those works have systematically investi- gated the impact of the many parameters controlling their approach. The present study aims at doing just this. As a test- case, we address the task of translating terms of the medical domain by exploit- ing pages mined from Wikipedia. One in- teresting outcome of this study is that sig- nificant gains can be obtained by using an association measure that is rarely used in practice. 1 Introduction Identifying translations of terms in comparable corpora is a challenge that has attracted many re- searchers. A popular idea that emerged for solv- ing this problem is based on the assumption that the context of a term and its translation share sim- ilarities that can be used to rank translation candi- dates (Fung, 1998; Rapp, 1999). Many variants of this idea have been implemented. While a few studies have investigated pattern matching approaches to compare source and tar- get contexts (Fung, 1995; Diab and Finch, 2000; Yu and Tsujii, 2009), most variants make use of a bilingual lexicon in order to translate the words of the context of a term (often called seed words). D´ejean et al. (2005) instead use a bilingual the- saurus for translating these. Another distinction between approaches lies in the way the context is defined. The most com- mon practice, the so-called window-based ap- proach, defines the context words as those cooc- curing significantly with the source term within windows centered around the term.1 Some studies have reported gains by considering syntactically motivated co-occurrences. Yu and Tsujii (2009) propose a resource-intensive strategy which re- quires both source and target dependency parsers, while Otero (2007) investigates a lighter approach where a few hand coded regular expressions based on POS tags simulate source parsing. The latter approach only requires a POS tagger of the source and the target languages as well as a small par- allel corpus in order to project the source regular expressions. Naturally, studies differ in the way each co- occurrence (either window or syntax-based) is weighted, and a plethora of association scores have been investigated and compared, the like- lihood score (Dunning, 1993) being among the most popular. Also, different similarity measures have been proposed for ranking target context vec- tors, among which the popular cosine measure. The goal of the different authors who inves- tigate context-projection approaches also varies. Some studies are tackling the problem of iden- tifying the translation of general words (Rapp, 1999; Otero, 2007; Yu and Tsujii, 2009) while others are addressing the translation of domain specific terms. …

Content truncated to 500 words