A Contextual Language Model to Improve Machine Translation of Pronouns by Re-ranking Translation Hypotheses

This paper addresses the translation divergencies of pronouns from English to French, specifically it and they, which have several gendered and non-gendered possible translations into French. Instead of using anaphora resolution, which is error-prone, we build a target language model that estimates the probabilities of a tuple of consecutive nouns followed by a pronoun. We bring evidence for the linguistic validity of the model, showing that the probability of observing a pronoun with a given gender and number increases with the proportion of nouns with the same gender and number preceding it. We use this French language model to re-rank the translation hypotheses generated by a phrase-based statistical machine translation system. While none of the pronoun-focused translation systems at the DiscoMT 2015 shared task improved over the baseline, our proposal achieves a modest but statistically significant improvement over it.