Document level Relationship Extraction based on context feature enhancement

Abstract

Document level Relationship Extraction (DocRE) tasks aim to extract relationships between multiple entities from long texts. However, obtaining feature representations for entity pairs that span multiple sentences is a challenge. Additionally, the feature information for triplets depends on both intra-document and inter-sentence information. To address this issue, this paper proposes a model named Plus-DocRE for DocRE(PDRE). Firstly, we introduce entity segmentation based on spans to increase the potential number of entities and improve negative sample recognition. Secondly, we utilize the BERT pre-trained model to obtain paragraph and local context information, enriching the features of entity pairs. Finally, through linear layers and self-attention mechanisms, we fuse the features of local and paragraph context for multi-label relationship classification, enabling entity relationship extraction. Meanwhile, we introduce a new data mechanism, C-DocRE, to simulate a more realistic scenario with annotation errors. Experimental results show that the PDRE model outperforms other baseline models in performance, achieving an F1 score of 53.6.