Exploiting Legal Event Detection Based on the BERT and Graph Convolutional Network

In a vast amount of legal documents and written judgement, there existing numerous dynamic processes of crime events, which serve as the basis for judges to understand the case and make the fair legal judgments. Therefore, it is very important for courts or the other judicial organs to extract events from a large number of legal texts using automated methods. With the development of deep learning, many scholars have proposed numerous legal event detection models based on deep learning technologies. However, these methods typically first only utilize the semantic information of context which mainly uses BERT to output their representations, and then detect the legal event trigger in the form of sequence labeling, whereas their efforts ignore the overfitting issues inherent in BERT itself, which leads to performance bottlenecks in the designing event detection models. To address the above issues, this paper designs a legal event detection method based on Graph Convolutional Networks (GCN) which can achieve the re-examining of attention weights. Specifically, this method can first construct a fully connected adjacency matrix acquired from the multi-head attention weights of each transformer layer in BERT, and then aggregate the contextual token nodes based on the edge weights inside the adjacency matrix, and finally achieve the re-examination of the BERT attention weights. To validate the effectiveness of the proposed model, some evaluated experiments are conducted on the large-scale Chinese legal event detection dataset LEVEN. The final experimental results demonstrate that compared to traditional BERT-based methods, the proposed GCN-based method achieves better performance in the legal event detection task.