Evaluation techniques applied to domain tuning of MT lexicons

We describe a set of evaluation techniques applied to domain tuning of bilingual lexicons for machine translation. Our overall objective is to translate a domain-specific document in a foreign language (in this case, Chinese) to English. First, we perform an intrinsic evaluation of the effectiveness of our domain-tuning techniques by comparing our domain-tuned lexicon to a manually constructed domain-specific bilingual termlist. Our results indicate that we achieve 66% recall and 95% precision with respect to a human-derived gold standard. Next, an extrinsic evaluation demonstrates that our domain-tuned lexicon improves the Bleu scores 50% over a statistical system—with a smaller improvement when the system is trained on a uniformly-weighted dictionary.