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

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 512–520, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics An MT Error-driven Discriminative Word Lexicon using Sentence Structure Features Jan Niehues and Alex Waibel Institute for Anthropomatics Karlsruhe Institute of Technology, Germany firstname.secondname@kit.edu Abstract The Discriminative Word Lexicon (DWL) is a maximum-entropy model that pre- dicts the target word probability given the source sentence words. We present two ways to extend a DWL to improve its abil- ity to model the word translation probabil- ity 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 tar- get sentence. We propose to include this structure by modeling the source sentence as a bag-of-n-grams and features depend- ing on the surrounding target words. Fur- thermore, as the standard DWL does not get any feedback from the MT system, we change the DWL training process to ex- plicitly focus on addressing MT errors. By using these methods we are able to im- prove the translation performance by up to 0.8 BLEU points compared to a system that uses a standard DWL. 1 Introduction In many state-of-the-art SMT systems, the phrase- based (Koehn et al., 2003) approach is used. In this approach, instead of building the translation by translating word by word, sequences of source and target words, so-called phrase pairs, are used as the basic translation unit. A table of correspon- dences between source and target phrases forms the translation model. Target language fluency is modeled by a language model storing monolingual n-gram occurrences. A log-linear combination of these main models as well as additional features is used to score the different translation hypotheses. Then the decoder searches for the translation with the highest score. One problem of this approach is that bilingual context is only modeled within the phrase pairs. Therefore, different approaches to increase the context available during decoding have been pre- sented (Haque et al., 2011; Niehues et al., 2011; Mauser et al., 2009). One promising approach is the Discriminative Word Lexicon (DWL). In this approach, a discriminative model is used to predict the probability of a target word given the words in the source sentence. In contrast to other models in the phrase-based system, this approach is capable of modeling the translation probability using information from the whole sentence. Thus it is possible to model long-distance dependencies. But the model is not able to use the structure of the sentence, since the source sentence is modeled only as a bag- of-words. Furthermore, the DWL is trained to discriminate between all translation options with- out knowledge about the other models used in a phrase-based machine translation system such as the translation model, language model etc. In contrast, we try to feedback information about possible errors of the MT system into the DWL. Thereby, the DWLs are able to focus on improving the errors of the other models of an MT system. We will …

Content truncated to 500 words