A Topic-Triggered Translation Model for Statistical Machine Translation

Translation model containing translation rules with probabilities plays a crucial role in statistical machine translation. Conventional method estimates translation probabilities with only the consideration of cooccurrence frequencies of bilingual translation units, while ignoring document-level context information. In this paper, we extend the conventional translation model to a topic-triggered one. Specifically, we estimate topic-specific translation probabilities of translation rules by leveraging topical context information, and online score selected translation rules according to topic posterior distributions of translated sentences. As compared with the conventional model, our model allows for more fine-grained distinction among different translations. Experiment results on large data set demonstrate the effectiveness of our model.