Entity–relation triple extraction based on relation sequence information

Abstract

Data overlap is a significant challenge in the task of entity–relation triple extraction. This task includes two research lines, line one first identifies entities and then predicts relations while line two completely shuffles the order. The methods in line two are more conducive to the optimization of the data overlap problem. Recent works have made breakthroughs in dealing with overlapping data, but there are still some defects such as difficulty in convergence and poor performance on datasets with numerous relations. To solve the above problems, we adopt a two-step strategy of first extracting subjects, and then predicting relation–object pairs. Considering the absence of connectivity between the two steps in the conventional method, we adopt the relation sequence as input in both steps and propose the TERS model. The relation sequence can connect two steps and improve the single-step and comprehensive extraction capability of the model. The TERS model consists of two modules. The first module performs information interaction and filters invalid subjects through the Text Relation Attention method. The second module implements multiple iterations of information interaction through the Information Flow method. The combination of these two modules contributes to the strong entity–relation triple extraction capability of our model. We evaluate our method on three public datasets. Extensive experiments show that our TERS model outperforms previous state-of-the-art models in triple extraction and overlapping data processing. Compared with other two-step extraction models, the advantages of our model are more obvious.