Pre-classification Supporting Reasoning for Document-level Relation Extraction

The document-level relation extraction task aims to extract relational triples from a document consisting of multiple sentences. Most previous models focus on modeling the dependency between the entities and neglect the reasoning mechanism. Some other models construct paths implicitly between co-sentence entities to find semantic relations. However, they ignore that there are interactions between different triples, especially some triples play an import role in predicting others. In this short research paper, we propose a new two stage framework PCSR(Pre-classification Supporting Reasoning) which captures the interactions between triples and utilizes these information for reasoning. Specifically, we make a pre-classification for each entity pair in the first stage. Then we aggregate the embeddings of predicted triples to enhance entity representation and make a new classification. Since the second classification could find triples missed in the first stage, we take the result as the supplement of the prior one. Experiments on DocRED show that our method achieves an F1 score of 62.11. Compared with the previous state-of-the-art model, our model increase by 0.81 on the test set, which demonstrates the effectiveness of our reasoning mechanism.