Modeling relation paths for knowledge base completion via joint adversarial training

Highlights：

• This paper proposes a novel adversarial method to model the relation paths between entity pair, which can extract the shared information between single relation (1-hop path) and multi-hop paths.

• This paper uses hierarchical attention networks to encode multi-hop paths between entity pair, which has been verified can extract valuable relations in the multi-hop path.

• Our model can be utilized to implement knowledge base completion, which achieves state-of-the-art experiment results.

• Each sub-module in our model can be interpreted well through experimental outputs.

• Our model can be generalized to many similar tasks, e.g., relation extraction.

摘要

•This paper proposes a novel adversarial method to model the relation paths between entity pair, which can extract the shared information between single relation (1-hop path) and multi-hop paths.•This paper uses hierarchical attention networks to encode multi-hop paths between entity pair, which has been verified can extract valuable relations in the multi-hop path.•Our model can be utilized to implement knowledge base completion, which achieves state-of-the-art experiment results.•Each sub-module in our model can be interpreted well through experimental outputs.•Our model can be generalized to many similar tasks, e.g., relation extraction.

论文评审过程：Received 3 June 2019, Revised 22 March 2020, Accepted 1 April 2020, Available online 7 May 2020, Version of Record 22 May 2020.