Semantic and syntactic information for neural machine translation

Abstract Introducing factors such as linguistic features has long been proposed in machine translation to improve the quality of translations. More recently, factored proven still be useful case sequence-to-sequence systems. In this work, we investigate whether gains hold state-of-the-art architecture neural translation, Transformer, instead recurrent architectures. We propose a new model, Factored introduce an arbitrary number word source sequence attentional system. Specifically, suggest two variants depending on level at which are injected. Moreover, combination mechanisms for and words themselves. experiment both with classical semantic extracted from linked data database, low-resource datasets. With best-found configuration, show improvements 0.8 BLEU over baseline Transformer IWSLT German-to-English task. more challenging FLoRes English-to-Nepali benchmark, includes very distant languages, obtain improvement 1.2 BLEU. These achieved not information.