Distantly Supervised Relation Extraction Based on Pre-trained Language Models and Dual-Modal Encoders

To solve the problems of insufficient semantic information representation in text and inadequate information transmission, leading to limited noise recognition capability and insufficient learning of long-tail relationships in distant supervised relation extraction, in this paper, a two-stage framework was proposed to integrate a pre-trained model (BERT) into multi-instance learning. Firstly, a pre-trained language model was utilized to learn text semantics so as to identify and mitigate noise. And than, a dual-modal encoder was designed within the framework to automatically learn the propagation patterns of entity types and relationships, tackling the long-tail problem. Experimental results on two widely-used datasets, NYT-10 and GDS, demonstrate that the proposed method can achieve significant improvements in both noise reduction and long-tail relation extraction.