Dynamic Models in Moses for Online Adaptation

Abstract

Abstract A very hot issue for research and industry is how to effectively integrate machine translation (MT) within computer assisted translation (CAT) software. This paper focuses on this issue, and more generally how to dynamically adapt phrase-based statistical machine translation (SMT) by exploiting external knowledge, like the post-editions from professional translators. We present an enhancement of the Moses SMT toolkit dynamically adaptable to external information, which becomes available during the translation process, and which can depend on the previously translated text. We have equipped Moses with two new elements: a new phrase table implementation and a new LM-like feature. Both the phrase table and the LM-like feature can be dynamically modified by adding and removing entries and re-scoring them according to a time-decaying scoring function. The final goal of these two dynamically adaptable features is twofold: to create additional translation alternatives and to reward those which are composed of entries previously inserted therein. The implemented dynamic system is highly configurable, flexible and applicable to many tasks, like for instance online MT adaptation, interactive MT, and context-aware MT. When exploited in a real-world CAT scenario where online adaptation is applied to repetitive texts, it has proven itself very effective in improving translation quality and reducing post-editing effort.