DPFSI: A legal judgment prediction method based on deontic logic prompt and fusion of law article statistical information

Abstract

The application of large language models in legal text processing faces challenges due to a lack of specialized legal knowledge and difficulty in handling complex legal logic and long texts. To address these issues, this paper introduces a legal judgment prediction method that combines deontic prompt learning, multi-granularity graph representation fusion, and law article statistical information integration. The model first modifies a pre-trained language model through deontic prompt learning, enhancing the semantic richness and interpretability of word vectors in structured texts. Additionally, it constructs multiple heterogeneous graphs to capture semantic and syntactic relationships, and applies a Gated Graph Neural Network for multi-granularity graph representation fusion. This fusion provides a comprehensive understanding of legal texts by effectively integrating semantic and structural information. Furthermore, the model calculates word frequency statistics between law articles and text sequences, using a $\beta$-variational autoencoder to align and integrate statistical information into text sequence features. This adaptive fusion reduces overfitting by replacing low-confidence neurons with global statistical information. Experimental results on the CAIL2018 dataset show that our approach achieves a significant improvement over existing methods, particularly in law article prediction, demonstrating its effectiveness and potential for enhancing legal AI applications.