French-English Terminology Extraction from Comparable Corpora

French-English Terminology Extraction from Comparable Corpora B´eatrice Daille and Emmanuel Morin University of Nantes, LINA - FRE CNRS 2729, 2, rue de la Houssini`ere - BP 92208, 44322 Nantes Cedex 3, France {beatrice.daille, emmanuel.morin}@univ-nantes.fr Abstract. This article presents a method of extracting bilingual lexica composed of single-word terms (SWTs) and multi-word terms (MWTs) from comparable corpora of a technical domain. First, this method ex- tracts MWTs in each language, and then uses statistical methods to align single words and MWTs by exploiting the term contexts. After ex- plaining the difficulties involved in aligning MWTs and specifying our approach, we show the adopted process for bilingual terminology ex- traction and the resources used in our experiments. Finally, we evaluate our approach and demonstrate its significance, particularly in relation to non-compositional MWT alignment. 1 Introduction Traditional research into the automatic compilation of bilingual dictionaries from corpora exploits parallel texts, i.e. a text and its translation [17]. From sentence- to-sentence aligned corpora, symbolic [2], statistical [11], or combined [7] tech- niques are used for word and expression alignments. The use of parallel corpora raises two problems: – as a parallel corpus is a pair of translated texts, the vocabulary appearing in the translated text is highly influenced by the source text, especially for technical domains; – such corpora are difficult to obtain for paired languages not involving English. New methods try to exploit comparable corpora: texts that are of the same text type and on the same subject without a source text-target text relationship. The main studies concentrate on finding in such corpora translation candidates for one-item words. For example, the French SWT manteau is translated in English by mantle in the domain of forestry, shield in the domain of marine activities, and by coat in the domain of clothing. The method is based on lexical context analysis and relies on the simple observation that a word and its translation tend to appear in the same lexical contexts. Thus, for our three possible translations of manteau, three different lexical contexts are encountered which are expressed below by English lexical units: R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 707–718, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 708 B. Daille and E. Morin – manteau/mantle : vegetation, forest, wood. . . – manteau/shield : boat, sea, shipbuilding. . . – manteau/coat : cloth, cold, wear. . . These contexts can be represented by vectors, and each vector element represents a word which occurs within the window of the word to be translated. Translation is obtained by comparing the source context vector to each translation candidate vector after having translated each element of the source vector with a general dictionary. This method is known as the “direct context-vector approach”. Using this method, [10] extracts English-Chinese one-item candidate translations from two years of English and Chinese newspaper articles by matching the context vector with 76% precision on the first 20 candidates. From English-German newspaper corpora of 85 million words, [14] improves the precision to …

Content truncated to 500 words