Dynamic Topic Adaptation for Phrase-based MT

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 328–337, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Dynamic Topic Adaptation for Phrase-based MT Eva Hasler1, Phil Blunsom2, Philipp Koehn1, Barry Haddow1 1School of Informatics, University of Edinburgh 2Dept. of Computer Science, University of Oxford Abstract Translating text from diverse sources poses a challenge to current machine translation systems which are rarely adapted to structure beyond corpus level. We explore topic adaptation on a diverse data set and present a new bilingual vari- ant of Latent Dirichlet Allocation to com- pute topic-adapted, probabilistic phrase translation features. We dynamically in- fer document-specific translation proba- bilities for test sets of unknown origin, thereby capturing the effects of document context on phrase translations. We show gains of up to 1.26 BLEU over the base- line and 1.04 over a domain adaptation benchmark. We further provide an anal- ysis of the domain-specific data and show additive gains of our model in combination with other types of topic-adapted features. 1 Introduction In statistical machine translation (SMT), there has been a lot of interest in trying to incorporate in- formation about the provenance of training exam- ples in order to improve translations for specific target domains. A popular approach are mixture models (Foster and Kuhn, 2007) where each com- ponent contains data from a specific genre or do- main. Mixture models can be trained for cross- domain adaption when the target domain is known or for dynamic adaptation when the target domain is inferred from the source text under translation. More recent domain adaptation methods employ corpus or instance weights to promote relevant training examples (Matsoukas et al., 2009; Fos- ter et al., 2010) or do more radical data selection based on language model perplexity (Axelrod et al., 2011). In this work, we are interested in the dynamic adaptation case, which is challenging be- cause we cannot tune our model towards any spe- cific domain. In previous literature, domains have often been loosely defined in terms of corpora, for exam- ple, news texts would be defined as belonging to the news domain, ignoring the specific content of news documents. It is often assumed that the data within a domain is homogeneous in terms of style and vocabulary, though that is not always true in practice. The term topic on the other hand can describe the thematic content of a document (e.g. politics, economy, medicine) or a latent cluster in a topic model. Topic modelling for machine transla- tion aims to find a match between thematic context and topic clusters. We view topic adaptation as fine-grained domain adaptation with the implicit assumption that there can be multiple distributions over translations within the same data set. If these distributions overlap, then we expect topic adapta- tion to help separate them and yield better trans- lations than an unadapted system. Topics can be of varying granularity and are therefore a flexi- ble means to structure data that is not uniform …

Content truncated to 500 words