Extracting and Selecting Relevant Corpora for Domain Adaptation in MT

The paper presents scheme for doing Domain Adaptation for multiple domains simultaneously. The proposed method segments a large corpus into various parts using self-organizing maps (SOMs). After a SOM is drawn over the documents, an agglomerative clustering algorithm determines how many clusters the text collection comprised. This means that the clustering process is unsupervised, although choices are made about cut-offs for the document representations used in the SOM. Language models aren then built over these clusters, and used as features while decoding a Statistical Machine Translation system. For each input document the appropriate auxiliary Language Model most fitting for the domain is chosen according to a perplexity criterion, providing an additional feature in the log-linear model used by Moses. In this way, a corpus induced by an unsupervised method is implemented in a machine translation pipeline, boosting overall performance in an end-to-end experiment.