Improving Compositional Translation with Comparable Corpora

We improved the compositional term translation method by using comparable corpora. A bilingual lexicon consisting of pairs of word sequences within terms and their correlations is derived from a bilingual document-aligned corpus. Then, for an input term, compositional translations are produced together with their confidence scores by consulting the corpus-derived bilingual lexicon. Thus, we can select the correct translation for the input term from among as many candidate ones as possible. An experiment with a comparable corpus of Japanese and English scientific-paper abstracts demonstrated that compositional translation using the corpus-derived bilingual lexicon outperforms that using an ordinary bilingual lexicon. Future work includes the incremental improvement of the bilingual lexicon with correlations, the refinement of the confidence score, and the extension of the compositional translation model to allow word order to be changed.