A Semi-supervised Joint Entity and Relation Extraction Model Based on Tagging Scheme and Information Gain

Abstract

Joint entity and relation extraction, completing the entity recognition and relation extraction simultaneously, can better integrate the information between two tasks and reduce the errors of each task. The methods based on tagging scheme treat the joint extraction as a sequence labeling task and have achieved outstanding results. However, those tag-based methods are insufficient in making full use of the information between entities and relations. It maybe the reason why they show a relatively poor precision. In this paper, we propose a novel semi-supervised approach that combines the tagging scheme with an information gain module. The information gain module is a combination of distant supervision and attention mechanism, which is used for obtaining prior information of candidate entities and relations. We believe that it is important for the joint extraction results to make much of the links among them. Our tagging scheme adds distant supervision to tag entity words and relational words in the given sentences, and combines with attention mechanism to improve their weight. It effectively helps our bias objective function improve model performance. Experiments on public dataset show that our methods are better than most of the existing pipelined and joint learning methods.