Introduction
Legal Judgment Prediction (LJP) is the task of automatically determining legal outcomes based on case facts. This process typically involves three subtasks: predicting applicable law articles, specific charges, and the term of penalty. Despite its importance, existing models often struggle with two key issues: the inefficient utilization of task dependencies and a lack of sensitivity to "confusing" cases where small numerical details (e.g., the quantity of a smuggled item) drastically alter the penalty.
This paper introduces the Multi-Perspective Bi-Feedback Network (MPBFN), a novel framework designed to mimic the hierarchical and iterative reasoning of human judges. By incorporating a bidirectional feedback loop and a specialized Word Collocation Attention (WCA) mechanism, the model achieves state-of-the-art performance across all LJP subtasks.
Methodology
Multi-Perspective Bi-Feedback Network (MPBFN)
The core of the proposed approach is the MPBFN architecture, which moves beyond simple multi-task learning by implementing a sophisticated feedback system. Unlike traditional pipeline models that only flow information forward, MPBFN utilizes both forward and backward dependencies to refine predictions.
The architecture consists of three main components:
- Forward Prediction (FP): Predictions from "pre-order" tasks (e.g., Law) are mapped into latent states and merged with the case's semantic vector to inform "follow-up" tasks (e.g., Charge).
- Backward Verification (BV): Follow-up task results are used to "check" and refine the rationality of earlier predictions through a gating mechanism.
- Multi-Perspective Integration: The final output for each task integrates its initial prediction, forward-flowing context, and backward-flowing verification, ensuring a robust and self-consistent result.
Word Collocation Attention (WCA)
To address the challenge of numerical sensitivity, the authors developed the WCA mechanism. Standard NLP models often treat numbers as arbitrary tokens, ignoring their magnitude. WCA employs Numerical Embedding, where digits are embedded separately to preserve magnitude information. Furthermore, it uses the Stanford CoreNLP toolkit to identify Word Collocations (e.g., "stole" + "5000 USD"). An LSTM-based attention layer then weighs these collocations, ensuring the model captures the critical relationship between actions and quantities.
Experimental Evaluation
The model was evaluated using two datasets from the CAIL2018 (Chinese AI and Law challenge): CAIL-small (~100k cases) and CAIL-big (~1.5M cases). Preprocessing involved filtering for single-label cases and removing charges with fewer than 100 occurrences to ensure statistical reliability.
Key Results
The MPBFN-WCA model consistently outperformed all baselines, including the previous state-of-the-art, TOPJUDGE.
The most significant improvements were observed in the Term of Penalty prediction. On the large-scale CAIL-big dataset, the model achieved a 13.3% improvement in F1-score over TOPJUDGE. Ablation studies confirmed that both the Bi-Feedback (BV) and Word Collocation (WCA) modules are essential, with WCA playing a pivotal role in the difficult task of determining prison terms.
Conclusion
The MPBFN-WCA framework demonstrates that legal judgment prediction is most effective when the model architecture reflects the logical topology of the legal domain. By combining bidirectional task verification with magnitude-aware numerical processing, the model provides a more accurate and human-like approach to LJP. Future work will focus on extending the model to handle multi-label charges and multi-defendant scenarios.
Summary updated