GCN2-NAA: Two-stage Graph Convolutional Networks with Node-Aware Attention for Joint Entity and Relation Extraction
Weicai Niu, Quan Chen, Weiwen Zhang, Jianwen Ma, Zhongqiang Hu · 2021
Joint extraction of entities and relations is critical for many tasks of Natural Language Processing (NLP), which aims to extract all triplets in the text. However, the huge challenge is that a sentence usually contains overlapping triplets. In this paper, we propose a joint extraction framework named GCN2-NAA based on a two-stage Graph Convolutional Neural networks (GCN) and Node-Aware Attention mechanism. We obtain multi-granularity representations and regional features of words by stacking multiple feature encoders and 1st-phase GCN. Besides, the node-aware attention mechanism and 2nd-phase GCN to capture the soft attention correlation matrix between all words in each relation type. Based on the constructed soft attention correlation matrix, we utilize GCN to further obtain the interaction between entities, relations, and triplets. Experiment results show that GCN2-NAA outperforms baseline models by 6.5% and 11.4% in terms of F1 score on NYT and WebNLG datasets, respectively.