Thursday, August 05, 2010
10:00 AM
Computer Science Bldg. Room 601
Ding Lui
University of Rochester
Towards Semantic-based Statistical Machine Translation
Syntax-based statistical machine translation (MT) systems are superior to the phrase-based MT systems in that they use tree-to-string templates to model long distance re-ordering in between two languages and generate more fluent sentences. To strengthen the syntax-based MT systems with semantic features, we propose a conditional log-linear model for learning the weights of the semantic-role features as well as the basic tree-to-string templates. We show significant improvement (over the EM algorithm) of the BLEU score by using the log-linear model, and significant improvement of sentences' fluency by adding the semantic role features, based on manual evaluation.
Automatic evaluation is crucial to the development of MT systems, especially when modern MT systems start using the evaluation metrics in the system tuning. We also propose a set of syntax-based features for MT evaluation, which show improvement in evaluating the sentence's fluency over the standard metric BLEU.