Re-ranking for Bilingual Lexicon Extraction with Bi-directional Linear Transformation from Comparable Corpora.

Recently a simple linear transformation with word embedding has been found to be highly effective to extract a bilingual lexicon from comparable corpora. However, the assumption that the pairs of bilingual word embedding for training this transformation satisfy a linear relationship automatically actually cant be guaranteed absolutely in practice. So the transformation of the source language to the target one is not consistent with the one of the target language to the source one. Given the translation candidate n-best list of a source word, we propose a bi-directional linear transformation based re-ranking method by combining the two direction linear score. The experimental results confirm that the proposed solution can achieve a significant improvement of 69% in the precision at Top-1 over the unidirectional baseline approach on the English-to-Chinese bilingual lexicon extraction task.