Learning an Expert from Human Annotations in Statistical Machine Translation: the Case of Out-of-Vocabulary Words