A Comparison of Smoothing Techniques for Bilingual Lexicon Extraction from Comparable Corpora

Smoothing is a central issue in language modeling and a prior step in different natural language processing (NLP) tasks. However, less attention has been given to it for bilingual lexicon extraction from comparable corpora. If a first work to improve the extraction of low frequency words showed significant improvement while using distance-based averaging (Pekar et al., 2006), no investigation of the many smoothing techniques has been carried out so far. In this paper, we present a study of some widelyused smoothing algorithms for language n-gram modeling (Laplace, Good-Turing, Kneser-Ney...). Our main contribution is to investigate how the different smoothing techniques affect the performance of the standard approach (Fung, 1998) traditionally used for bilingual lexicon extraction. We show that using smoothing as a preprocessing step of the standard approach increases its performance significantly.