Event Detection Method Based on Bidirectional Linear State Space

Aiming at the increasing granularity of event type and argument type classification in current event detection tasks, modules with more powerful feature extraction functions are needed. This paper proposed an event detection model based on BERT-BimamB-CRF. BiMamba is based on Mamba and incorporates the idea of bidirectional network, adding a backward SSM module to the Mamba module. The BERT pre-training model is used to analyze the character-level text, and the input text is mapped into a sequence of dynamic character embedding vectors. Then the bidirectional Mamba is used to fuse the context information and further extract the feature vector of the corresponding position. Finally, the CRF layer was used to calculate the optimal prediction path and output the final event element prediction result. This method is tested on Baidu’s open source event detection dataset DuEE1.0. The experimental results show that the BERT-BiMamba-CRF event detection model achieves good results in terms of accuracy, recall rate, F1 score and other indicators on the dataset, which is substantially better than the existing models.