Joint Extraction of Entities and Relations Based on Multi-label Classification

Abstract

Relations extraction is an important semantic analysis task in the field of natural language processing. We propose a new end-to-end model for joint extraction of entities and relations. In this model, we use a matching mechanism to match the relations and entities, transforming the joint extraction task into a combination of multi-label classification and sequence labeling task, and we use the dilated convolution combined with LSTM to encode words, more fully extract the semantics in the text. Firstly, we detect relations and entities separately, where we consider relations extraction as a multi-label classification task. Then we match the detected relations with the detected multiple entities. We conduct experiments on a dataset generated by distant supervision method. The accuracy of our experimental results reach 0.693, the recall rate is 0.430, and the F1 is 0.531, which exceed the experimental results of most models. The experimental results show that our method is better than most of the current joint extraction methods and can effectively extract overlapping relations.