中国科学院自动化研究所机构知识库

Neural machine translation (NMT) can achieve promising translation quality on resource-rich languages due to end-to-end learning. However, the widely-used NMT system only focuses on modeling the inner mapping from source to target without resorting to external knowledge. In this paper, we take English-Chinese translation as a case study to exploit the use of knowledge graph (KG) in NMT. The main idea is utilizing the entity relations in knowledge graph as constraints to enhance the connections between the source words and their translations. Specifically, we design two kinds of constraints. One is monolingual constraint that employs the entity relations in KG to augment the semantic representation of the source words. The other is bilingual constraint which enforces the entity relations between the source words to be shared by their translations. In this way, external knowledge can participate in the translation process and help to model semantic relationships between source and target words. Experimental results demonstrate that our method outperforms the state-of-the-art system.