Mining for Domain-specific Parallel Text from Wikipedia

Previous attempts in extracting parallel data from Wikipedia were restricted by the monotonicity constraint of the alignment algorithm used for matching possible candidates. This paper proposes a method for exploiting Wikipedia articles without worrying about the position of the sentences in the text. The algorithm ranks the candidate sentence pairs by means of a customized metric, which combines different similarity criteria. Moreover, we limit the search space to a specific topical domain, since our final goal is to use the extracted data in a domain-specific Statistical Machine Translation (SMT) setting. The precision estimates show that the extracted sentence pairs are clearly semantically equivalent. The SMT experiments, however, show that the extracted data is not refined enough to improve a strong in-domain SMT system. Nevertheless, it is good enough to boost the performance of an out-of-domain system trained on sizable amounts of data.