A Document-Level Trigger-Word-Free Multi-Event Extraction Method for Criminal Documents

Document Event Extraction (DEE) aims to extract event records from a given document. The DEE task faces two key challenges: parameter dispersion and the multi-event problem. To address these challenges, this paper proposes a trigger-word-free document-level event extraction method for criminal documents. The method utilizes deep learning technology to extract entities and sentence semantic information using pretrained language models. It then extracts dependent syntactic information in the document synchronously using graph neural networks. Finally, it encodes sentence-level and document-level features and employs the parameter path extension method with a memory mechanism to extract event records. The experimental data and evaluation indexes indicate that the method achieves better performance on the dataset, providing an effective solution for event extraction in legal criminal documents. This, in turn, helps to improve the efficiency and accuracy of judicial trials.