Revisiting Negation in Neural Machine Translation

In this paper, we evaluate the translation of negation both automatically and manually, in English–German (EN–DE) English– Chinese (EN–ZH). We show that ability neural machine (NMT) models to translate has improved with deeper more advanced networks, although performance varies between language pairs directions. The accuracy manual evaluation EN?DE, DE?EN, EN?ZH, ZH?EN is 95.7%, 94.8%, 93.4%, 91.7%, respectively. addition, under-translation most significant error type NMT, which contrasts diverse profile previously observed for statistical translation. To better understand root negation, study model’s information flow training data. While our analysis does not reveal any deficiencies could be used detect or fix find often rephrased during training, make it difficult model learn a reliable link source target negation. finally conduct intrinsic extrinsic probing tasks on showing NMT can distinguish non-negation tokens very well encode lot about hidden states but nevertheless leave room improvement.