Investigations on Translation Model Adaptation Using Monolingual Data

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 284–293, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Investigations on Translation Model Adaptation Using Monolingual Data Patrik Lambert, Holger Schwenk, Christophe Servan and Sadaf Abdul-Rauf LIUM, University of Le Mans 72085 Le Mans, France FirstName.LastName@lium.univ-lemans.fr Abstract Most of the freely available parallel data to train the translation model of a statistical ma- chine translation system comes from very spe- cific sources (European parliament, United Nations, etc). Therefore, there is increasing interest in methods to perform an adaptation of the translation model. A popular approach is based on unsupervised training, also called self-enhancing. Both only use monolingual data to adapt the translation model. In this pa- per we extend the previous work and provide new insight in the existing methods. We report results on the translation between French and English. Improvements of up to 0.5 BLEU were observed with respect to a very com- petitive baseline trained on more than 280M words of human translated parallel data. 1 Introduction Adaptation of a statistical machine translation sys- tem (SMT) is a topic of increasing interest during the last years. Statistical (n-gram) language models are used in many domains and several approaches to adapt such models were proposed in the literature, for instance in the framework of automatic speech recognition. Many of these approaches were suc- cessfully used to adapt the language model of an SMT system. On the other hand, it seems more chal- lenging to adapt the other components of an SMT system, namely the translation and reordering mod- els. In this work we consider the adaptation of the translation model of a phrase-based SMT system. While rule-based machine translation rely on rules and linguistic resources built for that purpose, SMT systems can be developed without the need of any language-specific expertise and are only based on bilingual sentence-aligned data (“bitexts”) and large monolingual texts. However, while monolin- gual data are usually available in large amounts and for a variety of tasks, bilingual texts are a sparse re- source for most language pairs. Current parallel corpora mostly come from one domain (proceedings of the Canadian or European Parliament, or of the United Nations). This is prob- lematic when SMT systems trained on such corpora are used for general translations, as the language jar- gon heavily used in these corpora is not appropriate for everyday life translations or translations in some other domain. This problem could be attacked by ei- ther searching for more in-domain training data, e.g. by exploring comparable corpora or the WEB, or by adapting the translation model to the task. In this work we consider translation model adaptation with- out using additional bilingual data. One can dis- tinguish two types of translation model adaptation: first, adding new source words or/and new transla- tions to the model; and second, modifying the prob- abilities of the existing model to better fit the topic of the task. These two directions are complementary and could be simultaneously applied. …

Content truncated to 500 words