LegalChainReasoner

This repository contains the code and data for the paper "LegalChainReasoner: Grounding Criminal Judicial Opinion Generation via Structured Legal Chains".

LegalChainReasoner is a framework designed to unify Legal Reasoning and Sentencing Prediction into a single generation process — Criminal Judicial Opinion Generation (CJOG). Statutory provisions are formalized as structured Legal Chains (Premise → Situation → Conclusion triplets) and injected into an LLM through a Chain-Aware Encoding module, so that sentencing decisions are grounded in and justified by the same legal analysis that produces the reasoning text.

---

📂 Repository Structure

.
├── LegalChainReasoner.py        # Main training and evaluation script (Source code)
├── requirements.txt
├── README.md
├── data/
│   ├── LAIC/                    # Main data from LAIC-2021
│   ├── PCCD/                    # Zero-shot evaluation data from PCCD
│   └── CAIL/                    # Zero-shot evaluation data from CAIL-2018
└── legal_chain/                 # Structured Legal Chains
    ├── robbery/                   
    │── fraud/                   
    └── ...

---

🛠 Requirements

The project is implemented using PyTorch and HuggingFace Transformers. To set up the environment, please run:

pip install -r requirements.txt

This demo code is built upon Llama-3.2-3B. Ensure you have access to HuggingFace Hub.

Key Dependencies:

• Python >= 3.8
• PyTorch (CUDA support needed)
• Transformers
• Peft (for LoRA fine-tuning)
• Thulac (for Chinese word segmentation)
• Rouge (for evaluation)

---

📊 Datasets

We utilize three datasets for experiments:

1. LAIC-2021: Derived from the Legal AI Challenge 2021, containing cases with factual descriptions, reasoning, and sentencing outcomes.
2. PCCD: A curated set of complex cases from the People’s Court Case Database for zero-shot generalization evaluation.
3. CAIL-2018: A subset of CAIL2018, used zero-shot for sentencing prediction only.

The processed datasets are located in the data/ directory. Each line in the data files is a JSON object mainly containing:

Field	Description
caseCause	The crime type (e.g., "robbery").
justice	The case facts (Fact).
opinion	The court's reasoning (Reasoning).
judge	The prison term in months (Sentencing).

---

🔗 Legal Chains

For each crime type we provide three files:

File	Purpose
chain.txt	Legal Chain in the form Premise -> Situation -> Conclusion.
<crime>_nodes.json	Node-level metadata.
<crime>.txt	Edge list describing the chain graph.

---

📥 Requesting Full Data & Resources

To keep this release compact, the following resources are demonstrated here and are available to researchers upon request:

• Full LAIC-2021 training set
• Full curated PCCD evaluation set
• Full curated CAIL2018 evaluation set
• Full Legal Chains covering all crime categories used in the paper

To request access, please contact the authors.

---

🚀 Usage

1. Preparation

Ensure that the Legal Chain folder contains the extracted chain files for the relevant crimes. The model uses these external knowledge bases to construct the Chain-Aware embeddings.

2. Training & Evaluation

The main script handles both training (with LoRA) and evaluation. To start the training process:

python LegalChainReasoner.py

3. Configuration

The training hyperparameters and file paths are currently configured within the main script. To change settings (e.g., batch size, learning rate, or device), please modify the if __name__ == '__main__': block inside LegalChainReasoner.py:

if __name__ == '__main__':
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") # Set GPU device
    epochs = 100                    # Number of training epochs
    batch_size = 1                  # Training batch size
    
    # ... ensure the file paths below match your data directory
    train = open("./data/LAIC/train_data.json", 'r', encoding='utf-8').readlines()
    test = open("./data/LAIC/test_data.json", 'r', encoding='utf-8').readlines()

Note that the script ships with the MAE / RMSE / ROUGE evaluation loop. The additional metrics reported in the paper (BLEU-1/2/N, BERTScore, GPTScore-pairwise, Legal-QA) are computed with the standard external tools described in the paper's Appendix.