Relation extraction based on multi-head self-attention and feature fusion

Abstract:
Relation extraction based on distant supervision has received extensive attention from researchers because of its ability to quickly and easily acquire large-scale datasets. However, most of the current studies do not consider the improvement of the model effect of multiple features. Considering the critical role of words and phrases in understanding sentence semantics, this paper combines these two features to propose a multi-feature fusion relation extraction method. Firstly, a multi-head self-attention mechanism is introduced to enhance important word and phrase features respectively, and then the information of these two parts is integrated by feature fusion to obtain the semantically enhanced sentence feature representation, and finally the relationship prediction is completed in units of packets. This paper conducts comparative experiments with multiple models on New York Times (NYT) data set, and proves that the fused features are conducive to deepening the model's understanding of semantics and can improve the effect of relation extraction.