Abstract
Legal Judgment Prediction (LJP) is a critical task in legal AI, aiming to automate the prediction of charges, law articles, and prison terms from factual descriptions. However, traditional NLP models often struggle with "confusing" crimes—cases that share similar factual patterns but distinct legal definitions. We introduce SEMDR (Semantic-Aware Dual Encoder Model), a framework that utilizes a three-level Legal Clue Tracing mechanism to improve LJP performance. By integrating lexicon filtering, contrastive sentence learning, and graph-based reasoning, SEMDR effectively denoises factual input and learns robust boundaries between similar legal concepts. Our experiments on the CAIL2018 dataset demonstrate state-of-the-art results, particularly in distinguishing confusing charges and handling low-frequency scenarios.
Introduction
Legal Judgment Prediction (LJP) involves the automated analysis of criminal fact descriptions to predict judicial outcomes. While standard deep learning models perform well on clear-cut cases, they frequently stumble when faced with "confusing" crimes, such as the subtle distinction between theft and robbery. These errors often stem from the model's inability to filter narrative noise or capture the specific semantic nuances required for legal differentiation. Existing models typically treat fact descriptions as monolithic text blocks, failing to isolate the core legal elements—motivation, action, and harm—that drive judicial decision-making.
Methodology
SEMDR addresses these challenges through a dual-encoder architecture and a hierarchical Legal Clue Tracing mechanism organized into three reasoning levels:
1. Lexicon-Tracing
To mitigate the impact of irrelevant narrative details, SEMDR employs a lexicon-tracing module. Instead of processing the entire raw text, the model filters descriptions to extract core legal elements: Motivation, Action, and Harm. This denoising step ensures the model focuses on the specific keywords (e.g., "coerced," "slashing," "fled") that define the legal nature of the act.
2. Sentence Representation Learning
The model utilizes a BERT-based encoder optimized via contrastive learning. By applying dropout noise to representations during training, the model is forced to learn more robust and uniform embeddings. This approach increases the model's sensitivity to subtle differences between similar sentences, preventing the "representation collapse" often seen in standard transformer models when dealing with similar factual patterns.
3. Multi-Fact Reasoning
SEMDR constructs a Legal Judgment Reasoning Graph where criminal facts and instrument labels (charges, articles) serve as nodes. Using a Graph Attention Network (GAT), the model propagates information across these nodes. This allows the system to "consult" related cases and labels, refining its predictions based on aggregated evidence rather than isolated text segments.
Datasets and Experimental Setup
The framework was evaluated using the CAIL2018 dataset, a large-scale Chinese legal judgment corpus. The researchers utilized three subsets:
- CAIL-small: 142,136 cases for initial training.
- CAIL-big: Over 1.7 million cases for large-scale validation.
- CAIL-mixed: A custom subset curated to specifically test low-frequency charges and confusing scenarios, such as high-frequency charges that are frequently misjudged by baseline models.
Results and Discussion
SEMDR achieved state-of-the-art (SOTA) performance across multiple benchmarks. On the CAIL2018 dataset, it outperformed existing models by over 2.21% in predicting confusing charges. For law article prediction on the CAIL-small subset, SEMDR exceeded the strong BERT-Crime baseline by 5.61% in Macro-Precision.
Embedding Uniformity
A key insight from the study is the importance of embedding space organization. Traditional BERT embeddings for similar charges often overlap, making classification difficult. SEMDR’s contrastive learning approach successfully "pushes" different charges apart in the vector space, creating clear boundaries.
Qualitative Analysis
In qualitative tests, SEMDR demonstrated superior reasoning in "hard sell" cases. For instance, in a robbery case involving coercion that standard BERT misclassified as fraud, SEMDR correctly identified the charge by focusing on the "force" elements highlighted through its attention and lexicon-tracing mechanisms.
Conclusion
SEMDR represents a significant advancement in Legal Judgment Prediction by shifting the focus from raw text processing to semantic clue tracing. The integration of lexicon-based denoising and graph-based reasoning allows the model to handle the complexities of legal logic more effectively than standard NLP architectures. Future work will explore the integration of explicit symbolic logic and the application of this framework to different legal systems, such as Common Law.
Summary updated