Incorporating Syntactic Uncertainty in Neural Machine Translation with Forest-to-Sequence Model

Previous work on utilizing parse trees of source sentence in Attentional Neural Machine Translation was promising. However, current models suffer from a major drawback: they use only 1-best parse tree which may lead to translation mistakes due to parsing errors. In this paper we propose a forest-to-sequence Attentional Neural Machine Translation model which uses a forest instead of the 1-best tree. Our method utilizes phrase structure of source sentence given by exponentially many grammar trees (compacted in a packed forest) to compensate parsing errors. Experiments on English to Japanese, German and Persian dataset demonsterate superiority of our method over the treeto-sequence and vanilla Attentional Neural Machine Translation models.