Automatic detection of translated text and its impact on machine translation

We investigate the possibility of automatically detecting whether a piece of text is an orig-inal or a translation. On a large parallel English-French corpus where reference infor-mation is available, we find that this is pos-sible with around 90 % accuracy. We fur-ther study the implication this has on Machine Translation performance. After separating our corpus according to translation direction, we train direction-specific phrase-based MT sys-tems and show that they yield improved trans-lation performance. This suggests that taking directionality into account when training SMT systems may have a significant effect on out-put quality.