Word Co-occurrence Counts Prediction for Bilingual Terminology Extraction from Comparable Corpora

International Joint Conference on Natural Language Processing, pages 1392–1400, Nagoya, Japan, 14-18 October 2013. Word Co-occurrence Counts Prediction for Bilingual Terminology Extraction from Comparable Corpora Amir Hazem and Emmanuel Morin Laboratoire d’Informatique de Nantes-Atlantique (LINA) Universit´e de Nantes, 44322 Nantes Cedex 3, France {Amir.Hazem, Emmanuel.Morin}@univ-nantes.fr Abstract Methods dealing with bilingual lexicon extraction from comparable corpora are often based on word co-occurrence ob- servation and are by essence more effec- tive when using large corpora. In most cases, specialized comparable corpora are of small size, and this particularity has a direct impact on bilingual terminology ex- traction results. In order to overcome in- sufficient data coverage and to make word co-occurrence statistics more reliable, we propose building a predictive model of word co-occurrence counts. We compare different predicting models with the tra- ditional Standard Approach (Fung, 1998) and show that once we have identified the best procedures, our method increases sig- nificantly the performance of extracting word translations from comparable cor- pora. 1 Introduction Using comparable corpora for bilingual lexicon extraction is becoming more and more a mat- ter of interest, especially because of the easier availability of this kind of corpora comparing to parallel ones. Many researchers proposed a va- riety of approaches (Fung, 1995; Rapp, 1999; Chiao and Zweigenbaum, 2002; D´ejean et al., 2002; Morin et al., 2007; Laroche and Langlais, 2010, among others). While different improve- ments were achieved, the starting point remains words’co-occurrences as they represent the ob- servable evidence that can be distilled from a cor- pus. Hence, frequency counts for word pairs often serve as a basis for distributional methods. The main assumption underlying bilingual lexicon ex- traction is: two words are more likely to be a trans- lation of each other if they share the same lexi- cal contexts (Fung, 1998). The most popular ap- proach named, the Standard Approach (Fung and Mckeown, 1997; Rapp, 1999), makes use of this assumption to perform bilingual lexicon extrac- tion. While good results on single word terms (SWTs) can be obtained from large corpora of sev- eral million words (80% for the top 10-20 (Fung and Mckeown, 1997), 91% accuracy for the top 3 (Cao and Li, 2002)). Results drop significantly using specialized small corpora (60% for the top 20 (Chiao and Zweigenbaum, 2002; D´ejean et al., 2002; Morin et al., 2007). The reliability of co-occurrence counts greatly relies on the amount of data. Clearly, the larger the training corpus, the more representative it is likely to be, and thus the more reliable the statis- tics of words. Therefore, the number and distri- bution of types in the available small sample are not reliable estimators (Evert and Baroni, 2007). This latter fact motivates the necessity of an alter- native to the unreliable counts especially when us- ing small specialized comparable corpora. Statis- tical NLP often deals in the prediction of variables ranging from text categories to linguistic struc- tures to novel utterances. If large specialized com- parable corpora are not available, one way to ap- proach this …

Content truncated to 500 words