Conditioning on Dependencies: Learning from Dependency Trees in Non-autoregressive Machine Translation

Abstract

We present an insertion-based translation model that uses linguistic dependencies to guide the order of sentence generation. This information is used at training time to motivate the model to favor certain insertions in a way such that words are more often inferred based on their dependencies. Evaluation results on the WMT 2020 German-English dataset show that our method achieves higher BLEU scores than unguided baseline models under the same training conditions. Our work emphasizes the relevance of using linguistic knowledge in neural machine translation.