关键词: 文档级, 关系抽取, 证据, 提及注意, 片段特征

Abstract:

Document-level Relationship Extraction （DocRE） aims at identifying all the relationships that exist between entity pairs in a document. Aiming at the problems of ineffective use of evidence sentences as well as document information， and multiple mentions of entities， a multi-feature fusion DocRE model named EMF （Evidence Multi-feature Fusion） was constructed based on evidence-enhanced contextual features. Firstly， entity types were added before and after entities， and relationship text features were associated with entity mentions to obtain relationship-specific entity features. Secondly， fragment representations were obtained through different convolutional kernels， and multi-granularity fragment-level features perceived by entity pairs were obtained through the attention mechanism. Meanwhile， contextual features highly correlated with the entity pairs were enhanced by using evidence distribution. Finally， the above features were fused for relationship classification， and during inference， the obtained evidence was composed into a pseudo-document and input into the classifier together with the original document for relationship classification. Experimental results on DocRED （Document-level Relation Extraction Dataset）， a DocRE dataset， show that when using BERT_base as the PLM encoder， compared with the state-of-the-art model EIDER （EvIDence-Enhanced DocRE）， the EMF model has the Ign F1 and F1 improved by 0.42 and 0.41 percentage points respectively， and the F1 reached 62.89%. It can be seen that the EMF model pays more attention to the parts that are related to entities and relationships， improves the extraction accuracy， and has a good interpretability.

Key words: document-level, relationship extraction, evidence, mention attention, fragment feature