A region-based hypergraph network for joint entity-relation extraction

Qian Wan,Luona Wei,Xinhai Chen,Jie Liu

DOI: https://doi.org/10.1016/j.knosys.2021.107298

2021-09-01

Abstract:<p>In joint entity and relation extraction, the input document is divided into multiple potential entity regions and context regions, where the characteristics of entities and their relations can often be reflected in the context. Therefore, an effective joint modeling method designed toward the features of different regions can lead to superior performance of joint entity and relation extraction. Previous works tend to implement in-depth modeling only for potential entity regions, ignoring the importance of contextual information for joint entity and relation extraction. In this paper, we propose a Region-based Hypergraph Network (RHGN) for joint entity and relation extraction. The RHGN introduces the concept of regional hypernodes for the first time, and proposes a cooperative method of GCN and BiLSTM to generate hypernodes for each region. Then, a region-based relation hypergraph is constructed for fairly and efficiently aggregate the features of all regions in the sentence. In order to initialize and update the features of the edges and hypernodes in the hypergraph, a Sequence-Enhanced Graph (SEG) unit is designed. Finally, we perform comparison experiments with existing competitive models on three public datasets: the CoNLL04, SciERC and ADE datasets. Experimental results demonstrate that our model achieves a significant improvement over the previous models on both entity recognition and relation extraction, and it also shows superior performance for dataset with nested entities. Extensive additional experiments further confirm the effectiveness of our approach.</p>

computer science, artificial intelligence

What problem does this paper attempt to address?

The paper aims to address the issue of joint modeling for Named Entity Recognition (NER) and Relation Extraction (RE). Specifically, existing methods often focus deeply on potential entity regions while neglecting the importance of contextual information. Therefore, the authors propose a Region-based Hypergraph Network (RHGN) for the joint task of entity and relation extraction. This model introduces the concept of region hypernodes for the first time and proposes a method that combines GCN and BiLSTM to generate hypernodes for each region. Then, a region-based relation hypergraph is constructed to fairly and efficiently aggregate features from all regions in the sentence. To initialize and update the edge and hypernode features in the hypergraph, a Sequence-Enhanced Graph (SEG) unit is designed. Experimental results show that the model achieves significant performance improvements on three public datasets (CoNLL04, SciERC, and ADE) in both entity recognition and relation extraction, especially excelling on datasets with nested entities.