Adaptive gated universal information extraction for Chinese legal texts

Constructing knowledge graphs in legal domains requires simultaneous extraction of entities and relations. To reduce repeated modeling in traditional approaches, we adopt the Universal Information Extraction (UIE) model as a foundation and propose an enhanced variant named Adaptive Gated Universal Information Extraction (AGUIE). This study develops a new decoder based on the Adaptive Focusing Gated Attention Unit (AFGAU). This unit enhances the standard Gated Attention Unit (GAU) by integrating two key components—learnable dynamic convolution and reset/update gating mechanisms. Moreover, the study employs a cross-pointer structure as the output layer to better identify information boundaries. To support this study, we construct a domain specific dataset for extracting key information from legal judgment documents. Systematic comparative analysis and ablation studies demonstrate that AGUIE achieves significant performance gains over baseline UIE, with an F1 score of 85.56% on our legal judgment documents dataset. Additionally, we evaluate the model’s generalization on public datasets such as ACE04, ACE05, and CoNLL04, covering both entity recognition and relation extraction tasks. Experimental results indicate that AGUIE demonstrates competitive results with recent studies on ACE04-Ent and CoNLL04, outperforms them on the ACE05 dataset, achieving F1 scores of 87.19% on ACE05-Ent and 79.29% on ACE05-Rel. In conclusion, AGUIE is a reliable and effective solution for universal information extraction in both legal and general domains.