Entity and Entity Type Enhanced Capsule Network for Distant Supervision Relation Extraction
Hongjun Heng, Renjie Li · 2021
In the task of distant supervision relation extraction, attention mechanism is introduced to distinguish correct instances from noise. However, the existing attention is often single-headed, and unable to recognize fine-grained word-level features from multiple subspaces. In addition, most current models do not fully utilize the external information of entities, and lack effective use of potential relation between the pre-trained entities. They only encode entities and other components of sentence into the vector, which is mixed with noise also weakens the potential relation. Therefore, we propose a dynamic double multi-head attention to filter out noise, and an entity supervisor to enhance the potential relation, and use capsule network to solve multi-label classification problem. We also propose auxiliary BGRU (Bidirectional Gated Recurrent Unit) to improve feature extraction performance. Experimental results show that our model achieves state-of-the-art performance on the precision-recall curve; Performance in processing of long-tail relations is also quite outstanding.