An Generative Entity Relation Extraction Model Based on UIE for Legal Text

Entity relation extraction is a basic technology in legal text analysis. The complexity of legal text description makes the traditional methods suffer from the problems of error propagation and low efficiency. Although generation based methods provide new solutions, domain-specific design is still necessary. We define ten entity types and three relation types for traffic accident crime cases. And then propose a new generative model UIE-ERNIE-CRF for chinese legal texts based on an universal information extraction model(UIE). Our model is divided into five layers, including input layer, a general semantic representation layer, a task semantic representation layer, a decoding layer, and the output layer. The model identifies specific semantic information required in different tasks through the ERNIE3.0 model. In order to solve the long-distance dependency problem faced by double pointer decoder in the UIE model, sequence labeling decoder CRF is introduced. Compared with the baseline model, UIE-ERNIE-CRF performs well on the precision rate, recall rate and F1 value. And the ablation experiment shows that introducing the ERNIE and CRF is effective for entity relation extraction of legal text.