Table 1 Comparative summary of recent studies on legal judgment prediction, highlighting datasets, methods, and evaluation metrics.
Ref | Method | Techniques | Dataset | Results | Prediction Task |
|---|---|---|---|---|---|
Text-CNN, RNN, Wide&TextCNN, TextDenseNet | Deep Learning Multi-Fusion Models | BDCI 2017 Chinese criminal cases | 83% Precision | Predict judgment outcomes: specifically, predict applicable law terms | |
SVM + LSTM + CNN Hybrid | Ensemble deep learning | ECtHR | 75% accuracy | Predict court case rulings from case facts. | |
Attention mechanism + LSTM | Knowledge-Injected Co-Attention (AutoJudge) | Private Loan case dataset | 81% F1-Score | Predict judgment outcomes for each claim | |
RF + BERT | Rationale-Based LJP (RLJP) | Real-world Chinese criminal cases | 71% Macro-F1 | Predicts the full legal judgment | |
Hier-BERT | HD-LJP (Hierarchical Dependency LJP) | CAIL | 74% Accuracy | Predicts multiple judgment aspects | |
LSTM + Knowledge tracing | Prompt Learning + Charge Keywords (Prompt4LJP) | BDCI 2017 dataset | 83% Accuracy | Predicts the charge (offense) of a case using prompt. | |
LSTM + CNN | LKEPL: Legal Knowledge-Enhanced Prompt Learning | CAIL2018 | 73% F1-Score | Predicts the full judgment from facts. | |
C-GNN | GCLA: Graph Contrastive Learning for LJP | Chinese legal datasets | 88% Recall | Predicts standard LJP outputs – applicable law article(s), charge, and penalty. | |
RoBERTa Large LLM | KnowJudge/KnowPrompt4LJP | CAIL2018 | 78% Accuracy | Predicts the charge/offense for a given case | |
LSTM + CNN | NumLJP: Numerical LJP with Magnitude Reasoning | criminal cases focusing on theft, fraud cases where prison term and/or fines | 56.09% Macro-F1 | Predicts numeric judgment outcomes | |
T5 Model | MS-Judge: Multi-Stage Case Representation | Civil cases | 85% Recall | Predicts judgment outcomes for each claim in a civil case | |
LSTM + BERT | NeurJudge: Circumstance-Aware Framework | CAIL2018 dataset | 73% Accuracy | Predicts the full judgment (law article, charge, penalty). | |
Lawformer (Legal Longformer) | Domain-specific long-document Transformer for law. | Chinese legal documents | 82% Precision | Supports legal judgment prediction tasks by encoding entire case documents. | |
BERT + Attention LLM | AgentsBench: Multi-Agent LLM Simulation | Synthetic and real case scenarios | 79% Accuracy | Simulates legal judgment prediction as a courtroom debate. | |
Hierarchical Attention + CNN | MHAN: Modified Hierarchical-Attention Network | High-court cases | ~ 78% accuracy | Predicts the case outcome given the case’s fact narrative | |
LSTM + RF + GRU | Enhanced Hybrid DL Model (HDLMSF) | ECtHR | 74% Accuracy | Predicts the final judgment outcome of cases binary |
