Video Recordings

Work on domain adaptation in machine translation has primarily focused on adapting model probabilities to correspond to the new domain. However, recent analyses suggest that for many domains, unknown words and new senses are the main sources of errors; different approaches to adaptation may therefore be appropriate. Our team at the CLSP workshop analyzed the problem of domain adaptation in more depth. We developed and evaluated several methods for discovering new senses and mining corpora for translations. A large part of our work concerned integrating the method of phrase sense disambiguation into the Moses decoder and developing domain adaptation techniques for this method.