Parallel Training Data Selection for Conversational Machine Translation

We describe data selection strategies for an English-French Skype conversation translation task without in-domain training data. Selection methods based on language modeling and text formality criterion are evaluated. Our main finding is that translating conversation transcripts turned out to not be as challenging as we expected: while translation quality is of course not perfect, a straightforward phrase-based system trained on movie subtitles yields high BLEU scores, and small improvements are obtained by using a simple heuristic to select more Skype-like examples.