A Simple Overlapping Relation Extraction Method Based on Dropout
Yiming Zhan, Zhao Li, Xiuhao Zhao, Chao Zhang, Tong Chen · 2022 International Joint Conference on Neural Networks (IJCNN) · 2022
The relation extraction (RE) task of knowledge acquisition aims to identify all specific relations of two entities. Prior works had proved the effectiveness of tagging methods, especially on the overlapping triple extraction problem. In this work, we introduce a simple relation extraction method (DropRel) based on dropout. We propose a Dropout-Normalization layer to generate different vectors for one token pair. Our model also provides a unified perspective for entity boundary and relation tagging tasks. We define a novel dependence relation for relation triple to integrate boundary into relation. The empirical experiments show that the DropRel model outperforms previous state-of-the-art methods on NYT and WEBNLG datasets. And our ablation study shows the DropRel without the dependence relation is better on the WEBNLG dataset.