Extending Statistical Machine Translation with Discriminative and Trigger-Based Lexicon Models

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 210–218, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Extending Statistical Machine Translation with Discriminative and Trigger-Based Lexicon Models Arne Mauser and Saˇsa Hasan and Hermann Ney Human Language Technology and Pattern Recognition Group Chair of Computer Science 6, RWTH Aachen University, Germany <surname>@cs.rwth-aachen.de Abstract In this work, we propose two extensions of standard word lexicons in statistical ma- chine translation: A discriminative word lexicon that uses sentence-level source in- formation to predict the target words and a trigger-based lexicon model that extends IBM model 1 with a second trigger, allow- ing for a more fine-grained lexical choice of target words. The models capture de- pendencies that go beyond the scope of conventional SMT models such as phrase- and language models. We show that the models improve translation quality by 1% in BLEU over a competitive baseline on a large-scale task. 1 Introduction Lexical dependencies modeled in standard phrase- based SMT are rather local. Even though the deci- sion about the best translation is made on sentence level, phrase models and word lexicons usually do not take context beyond the phrase boundaries into account. This is especially problematic since the average source phrase length used during decod- ing is small. When translating Chinese to English, e.g., it is typically close to only two words. The target language model is the only model that uses lexical context across phrase boundaries. It is a very important feature in the log-linear setup of today’s phrase-based decoders. However, its context is typically limited to three to six words and it is not informed about the source sentence. In the presented models, we explicitly take advan- tage of sentence-level dependencies including the source side and make non-local predictions for the target words. This is an important aspect when translating from languages like German and Chi- nese where long-distance dependencies are com- mon. In Chinese, for example, tenses are often en- coded by indicator words and particles whose po- sition is relatively free in the sentence. In German, prefixes of verbs can be moved over long distances towards the end of the sentence. In this work, we propose two models that can be categorized as extensions of standard word lex- icons: A discriminative word lexicon that uses global, i.e. sentence-level source information to predict the target words using a statistical classi- fier and a trigger-based lexicon model that extends the well-known IBM model 1 (Brown et al., 1993) with a second trigger, allowing for a more fine- grained lexical choice of target words. The log- linear framework of the discriminative word lexi- con offers a high degree of flexibility in the selec- tion of features. Other sources of information such as syntax or morphology can be easily integrated. The trigger-based lexicon model, or simply triplet model since it is based on word triplets, is not trained discriminatively but uses the classi- cal maximum likelihood approach (MLE) instead. We train the triplets iteratively on …

Content truncated to 500 words