Two Approaches to Correcting Homophone Confusions in a Hybrid Machine Translation System

In the context of a hybrid French-toEnglish SMT system for translating online forum posts, we present two methods for addressing the common problem of homophone confusions in colloquial written language. The first is based on hand-coded rules; the second on weighted graphs derived from a large-scale pronunciation resource, with weights trained from a small bicorpus of domain language. With automatic evaluation, the weighted graph method yields an improvement of about +0.63 BLEU points, while the rulebased method scores about the same as the baseline. On contrastive manual evaluation, both methods give highly significant improvements (p < 0.0001) and score about equally when compared against each other.