Detecting Cross-Lingual Semantic Divergence for Neural Machine Translation
Marine Jacinthe Carpuat, Yogarshi Vyas, Xing Hua Niu · 2017
Parallel corpora are often not as parallel as one might assume: non-literal translations and noisy translations abound, even in curated corpora routinely used for training and evaluation.We use a cross-lingual textual entailment system to distinguish sentence pairs that are parallel in meaning from those that are not, and show that filtering out divergent examples from training improves translation quality. IntroductionParallel sentence pairs provide examples of translation equivalence to train Machine Translation (MT) and cross-lingual Natural Language Processing.However, despite what the term "parallel" implies, the source and target language often do not convey the exact same meaning.This is a surprisingly common phenomenon, not only in noisy corpora automatically extracted from comparable collections, but also in parallel training and test corpora, as can be seen in Table 1.This issue has mostly been ignored in machine translation, where parallel sentences are assumed to be translations of each other, and translations are assumed to have the same meaning.Prior work on characterizing parallel sentences for MT has focused on data selection and weighting for domain adaptation (Foster and Kuhn, 2007; Axelrod et al., 2011, among others), and on assessing the relevance of parallel sentences by comparison with a corpus of interest.In contrast, we focus on detecting an intrinsic property of parallel sentence pairs.Divergent sentence pairs have been viewed as noise both in comparable and non-parallel corpora (Fung and Cheung, 2004;Munteanu and Marcu, 2005;AbduI-Rauf and Schwenk, 2009;Smith et al., 2010;Riesa and Marcu, 2012) and Divergent segments in OpenSubtitles en someone wanted to cook bratwurst.fr vous vouliez des saucisses grillées.gl you wanted some grilled sausages.en i don't know what i'm gonna do.fr j'en sais rien.gl i don't know.en -has the sake chilled?-no, it's fine.fr -c'est assez chaud?gl -it is hot enough?en you help me with zander and i helped you with joe.fr tu m'as aidée avec zander, je t'ai aidée avec joe.gl you helped me with zander, i helped you with joe. Divergent segments in newstest2012en i know they did.fr je le sais.gl i know it.en the female employee suffered from shock.fr les victimes ont survécu leur peur.gl the victims have survived their fear.