This paper describes the University of Edinburgh's submissions to the WMT17shared news translation and biomedical translation tasks. We participated in 12translation directions for news, translating between English and Czech, German,Latvian, Russian, Turkish and Chinese. For the biomedical task we submittedsystems for English to Czech, German, Polish and Romanian. Our systems areneural machine translation systems trained with Nematus, an attentionalencoder-decoder. We follow our setup from last year and build BPE-based modelswith parallel and back-translated monolingual training data. Novelties thisyear include the use of deep architectures, layer normalization, and morecompact models due to weight tying and improvements in BPE segmentations. Weperform extensive ablative experiments, reporting on the effectivenes of layernormalization, deep architectures, and different ensembling techniques.
translated by 谷歌翻译