Chinese Event Subject Extraction in the Financial Field Integrated with BIGRU and Multi-head Attention
Yangwenhao Liu, Changhui Liu, Liya Wang, Zhe Chen, Yuan Xin Wu · Journal of Physics Conference Series · 2021
Abstract Event subject extraction was to extract subjects of specific event types. For the traditional BiLSTM network, the threshold is complicated, the required parameters are many, and the time cost is high. This paper is oriented to the financial field and proposes a method of introducing a multi-head attention mechanism based on the BIGRU network to extract event subjects. First, the text is vectorized, and then the word vector obtained is input into the BIGRU network to learn the context features, and introduce a multi-head attention mechanism to extract the depth feature values of the text. Finally, a comparative experiment is conducted on the data set. The method in this paper achieves an Accuracy value of 80.47% and an F1 value of 89.18%. The result is better than the control group, indicating that the model proposed in this paper can effectively improve the accuracy of extracting Chinese event subjects.