Document-level Neural Machine Translation Using Dependency RST Structure

Anonymous

Abstract: Document-level machine translation (MT) extends the translation unit from the sentence to the whole document. Intuitively, discourse structure can be useful for document-level MT for its helpfulness in long-range dependency modelling. However, few efforts have been paid on leveraging discourse information for document-level neural machine translation(NMT). In this paper, we propose a dependency Rhetorical Structure Theory (RST) tree enhanced NMT model, RST-Transformer. The model only needs to encodes the dependency RST tree of the source document via the attention mask, and can enhance both the encoder and the decoder. Experiments on English-German datasets in both non-pretraining and pretraining settings show that our discourse information enhanced approach outperforms the current state-of-the-art document-level NMT model.

Paper Type: long