Extraction of translation equivalents from parallel corpora using sense-sensitive contexts

EAMT 2005 Conference Proceedings 97 Extraction of Translation Equivalents from Parallel Corpora Using Sense-sensitive Contexts Pablo Gamallo Otero Departamento de Língua Espanhola Universidade de Santiago de Compostela pablogam@usc.es Abstract. The paper proposes an unsupervised method to extract translation equivalents from parallel corpora. The strategy we use takes into account the context of words. Given a word of the source language and a particular context, we learn its word translation within an equivalent context. We first extract pairs of similar contexts and, then, we compare the similarity between words appearing in each pair. This allows us to use a very low threshold to identify correct translation equivalents. Moreover, as polysemic words tend to have dif- ferent senses in different context pairs, we are able to associate several translation equiva- lents to the same polysemic word. The main contribution of this paper is precisely to learn the correct translation equivalent of a word in a specific context. On the other hand, we do not align texts by detecting sentences or other small linguistic units. We identify natural boundaries by detecting explicit parts or segments of the corpus. Most text corpora contain natural boundaries to explicitly separate basic parts such as chapters, articles, receipts, legal documents, letters, etc. We use these explicit and natural parts to align parallel corpora. To compute similarity within these large segments, we define a particular version of the Dice coefficient. 1. Introduction Parallel corpora are a huge resevoir of bilingual lexical information. The extraction of transla- tion lexicons relies, to a certain extent, on paral- lel text alignment often to the sentence level. Most extraction methods use the co-occurrence frequencies and locations of expressions in al- igned sentences to compute the translation cor- relations between expression types in the two lan- guages (Gale and Church, 1991; Melamed, 1997; Ahrenberg and Andersson 1998; Tiedemann, 1998; Vintar, 2001; Kwong, 2004). These methods have at least two drawbacks: first, although a number of automatic sentence alignment meth- ods have been proposed, they are not very reli- able when corpora have unclear sentence bound- aries or with noisy bilingual texts (Fung and McKeown, 1996). Second, as these methods are mostly based on one-to-one word translations, they are not adapted to dealing with word poly- semia. To avoid the first drawback, we use a type of aligment based, not on sentence identifica- tion, but on identifying natural boundaries by de- tecting the explicit parts of the corpus. Most text corpora contain natural boundaries to explicitly separate basic parts such as chapters, articles, receipts, legal documents, letters, etc. We will use these explicit and natural parts to align par- allel corpora. Regarding the second drawback (i.e., polysemia), we will propose a method that takes into account word sense disambiguation in context. More precisely, the extraction me- thod we propose in this paper aims at identify- ing bilingual correspondences between sense-sen- sitive contexts. Equivalent sense-sensitive con- texts will allow us to extract correspondences between types of word senses. This is the main contribution of our …

Content truncated to 500 words