An MT Error-Driven Discriminative Word Lexicon using Sentence Structure Features

The Discriminative Word Lexicon (DWL) is a maximum-entropy model that predicts the target word probability given the source sentence words. We present two ways to extend a DWL to improve its ability to model the word translation probability in a phrase-based machine translation (PBMT) system. While DWLs are able to model the global source information, they ignore the structure of the source and target sentence. We propose to include this structure by modeling the source sentence as a bag-of-n-grams and features depending on the surrounding target words. Furthermore, as the standard DWL does not get any feedback from the MT system, we change the DWL training process to explicitly focus on addressing MT errors. By using these methods we are able to improve the translation performance by up to 0.8 BLEU points compared to a system that uses a standard DWL.