ICA for Bilingual Lexicon Extraction from Comparable Corpora

Independent component analysis (ICA) is a statistical method used to discover hidden features from a set of measurements or observed data so that the sources are maximally independent. This paper reports the first results on using ICA for the task of bilingual lexicon extraction from comparable corpora. We introduce two representations of data using ICA. The first one is called global ICA (GICA) used to design a global representation of a context according to all the target entries of the bilingual lexicon, the second one is called local ICA (LICA) and is used to capture local information according to target bilingual lexicon entries that only appear in the context vector of the candidate to translate. Then, we merge both GICA and LICA to obtain our final model (GLICA). The experiments are conducted on two different corpora. The French-English specialised corpus ’breast cancer’ of 1 million words and the French-English general corpus ’Le Monde / New-York Times’ of 10 million words. We show that the empirical results obtained with GLICA are competitive with the standard approach traditionally dedicated to this task.