Source Discriminative Word Lexicon for Translation Disambiguation
Teresa Herrmann, Jan Niehues, Alex Waibel · RWTH Publications (RWTH Aachen) · 2015
This paper presents a source discriminative word lexicon that performs translation disambiguation for individual source words using structural features, such as context and dependency relations in the sentence.The individual translation predictions are combined into sentence scores that are used in N -best list re-ranking to improve the translation output of a state of the art phrase-based machine translation system.The approach is used to improve explicitly the translation of word categories that require grammatical agreement to hold in the target language after translation, e.g.pronouns, as well as subjects and verbs.The results show that the translation predictions provided by the source discriminative word lexicon increase the prediction accuracy by up to 10%.The translation quality can be improved by up to 0.6 BLEU points on English-German translation.