Span-based Joint Extracting Subjects and Objects and Classifying Relations with Multi-head Self-attention

Entity recognition and relation extraction are two central tasks for extracting information from unstructured text. The current popular models are to perform joint entity and relationship extraction at the span level. However, they do not take full advantage of the uneven distribution of information values in natural languages and their essence of extracting entities before classifying relations leads to entity redundancy. To solve these problems, we propose a novel span-based joint entity and relationship extraction model. We embed the multi-head self-attention layer to the existing span-based entity and relation joint extraction model to leverage valuable information such as semantic and syntactic features in the input sentences. We also propose a novel method to extract subject and object, which can alleviate the entity redundancy problem. In the ablation experiments, we demonstrate the benefits of both improvements for extracting relation triples. The micro f1 of our model on Con1104 is 72.82, the macro f1 is 74.20, and the f1 on SciERC is 52.03, which are better than the currently widely used model.