Core Fidelity of Translation Options in Phrase-Based Machine Translation
Ondřej Hübsch · Digital Repository (National Repository of Grey Literature) · 2017
One popular approach to machine translation is to break sentences into small groups of contiguous words (phrases) and then to translate these phrases inde- pendently. Translations of these phrases are extracted beforehand from a large amount of bilingual data. The goal of this thesis is to detect semantical incorrect- ness in the extracted translations of phrases. One source of potential problems is poor quality of training data (high quality parallel data are very hard to ob- tain), more severe are possible problems introduced by double negative in Czech: the translated sentences might have a completely opposite meaning to the orig- inal one. We first tried to modify our prior work to penalize such erroneous translations. Then we designed and trained our own neural model to produce a semantical score for a given phrase translation. We evaluated the improvements on a small manually annotated set of translations and also in an end-to-end ma- chine translation task. Using our model in an end-to-end machine translation system yields a significant improvement of 0.5 BLEU over the baseline. Our model also beats an existing state-of-the-art method not only in the end-to-end translation (by 0.2 BLEU), but also on the manually annotated data by a factor of more than 2 in recognition of incorrect translations. 1