Computational Lexicography and Lexicology Elexbi, a Basic Tool for Bilingual Term Extraction from Spanish-Basque Parallel Corpora

We present the work done by Elhuyar Foundation in the field of bilingual terminology extraction. The aim ofthis work is to develop some techniques for the automatic extraction ofpairs ofequivalent terms from Spanish-Basque translation memories, and to implement those techniques in a prototype. Our approach is based on a monolingual extraction of term candidates in each language, then the creation of candidate bigrams from both segments of the same translation unit, and, finally, the selection of the most likely pair of candidates, mainly by the use of statistical information (association measures) and cognates. In the first step, we use linguistic techniques for the extraction of term candidates. The result of our work is ELexBI, a prototype tool that can extract equivalent terms from Spanish-Basque translation memories. This work wants to be a contribution to corpus-based bilingual lexicography and terminology in Basque. 1 Objective The aim of this work is to develop and apply techniques for the automatic extraction of pairs ofequivalent terms from Spanish-Basque translation memories, and to implement those techniques in a user-friendly prototype. This work can be located in a wider research area. In fact, the extraction of equivalent terms from translation memories can be seen as a particular case of the extraction of lexical equivalences from parallel corpora. In this first stage of development, the translation memories that we use as input are the product of translators' work; that is to say, the alignment at sentence level is 100% correct. In future work, automatically aligned memories and parallel corpora aligned at document level will be used. As for the type of term equivalents we attempt to find, we deal with one-word and multiword terms which have noun-phrase structure. Furthermore, equivalences between one-word and multiword terms are also taken into account. 2 Extraction process Different approaches have been proposed for the extraction of lexical correspondences from parallel corpora. Most of them are closely related to the task of word-level alignment. According to Tiedemann (2003) and Kraif (2002a, 2002b), we look at the extraction of lexical correspondences as a different but much related task.