Leveraging Attention Mechanism and Class Weighting for Legal Event Detection in Chinese Text

Event detection is an information extraction task that involves extracting specified event types from textual sequences. Currently, most event detection studies focus on English corpora; there is a lack of exploration in other linguistic contexts. Thus, a study on event detection in the Chinese corpus is essential. Sequence-based event detection has been extensively studied in the past, and many studies have utilized high-performance neural network models, such as traditional recurrent neural networks. This study aims to enhance the performance of sequence models by altering the base model, Bidirectional Long Short Term Memory (BiLSTM), to a Bidirectional Gated Recurrent Unit (BiGRU) and incorporating multi-head attention mechanisms, Conditional Random Fields (CRF), and class weights. These modifications not only improve the model’s accuracy but also enhance computational efficiency by reducing the number of parameters relative to large pre-trained models such as Bidirectional Encoder Representations from Transformers (BERT). The experimental findings demonstrate that the proposed model’s modification achieves an F1 Score of 83.55 for the micro standard and 78.07 for the macro standard. This presents a substantial improvement over the baseline, delivering performance nearly on par with state-of-the-art BERT-based models on the same dataset, while requiring significantly fewer parameters.