Source-linked AI summary
Learning to Predict Charges for Criminal Cases with Legal Basis
Bingfeng Luo, Yansong Feng, Jianbo Xu, Xiang Zhang, Dongyan Zhao
TL;DR
Charge prediction from case facts is difficult because charges may be subtle and multi-label, while users also need explicit legal bases. The paper jointly models charge prediction and relevant-article extraction with attention, and reports effective prediction, improved use of articles, and reasonable generalization across expression styles.
Problem
Charge prediction from fact descriptions must handle subtle distinctions and multiple crimes, while relevant law articles are needed as explicit legal basis for predictions.
Method
An attention-based neural network jointly models charge prediction and relevant-article extraction, using fact-guided attention to aggregate article evidence.
Results
The full model effectively predicts charges and extracts relevant articles, while experiments show that articles improve charge prediction and the model generalizes reasonably to news data.
Takeaways & Limitations
Relevant law articles provide legal basis and can improve charge prediction when integrated with fact descriptions and handled through attention.
Takeaways & Limitations
The dataset is restricted to one-defendant cases because matching facts and charges across defendants is hard and expensive.
Abstract
from arXiv · showhide
The charge prediction task is to determine appropriate charges for a given case, which is helpful for legal assistant systems where the user input is fact description. We argue that relevant law articles play an important role in this task, and therefore propose an attention-based neural network method to jointly model the charge prediction task and the relevant article extraction task in a unified framework. The experimental results show that, besides providing legal basis, the relevant articles can also clearly improve the charge prediction results, and our full model can effectively predict appropriate charges for cases with different expression styles.
1 Introduction
The paper frames charge prediction from case facts as a difficult multi-label task requiring subtle distinctions and legal-basis evidence. It proposes jointly modeling charge prediction and relevant-article extraction with attention, and reports effectiveness across expression styles.
- Motivation: Charge prediction assigns appropriate criminal charges from textual fact descriptions for legal assistant systems.Such systems can help users find similar cases, possible penalties, and legal bases despite difficult legal materials and terminology.
- Challenges: Subtle distinctions between charges, such as intentional homicide versus intentional injury, require inferring the defendant’s intent from facts.The example concerns whether the defendant intended to kill or only to hurt the victim.
- Challenges: A single case may involve multiple crimes, making charge prediction a multi-label classification problem.
- Legal basis: Relevant law articles should accompany predictions as explicit legal basis, especially in civil-law systems where judgments rely on statutory laws.Chinese judgments include relevant articles in the court-view part to support decisions.
- Prior work: Earlier approaches used shallow features, expert-designed factors, or simple classification paradigms, and treated charge prediction and article extraction independently.The paper identifies this separation as ignoring potential benefits between the related tasks.
- Approach: The proposed framework jointly models both tasks with sentence- and document-level Bi-GRUs and a two-stack attention mechanism.The method addresses latent correspondence from case facts to relevant articles and then to charges.
- Evidence: Experiments on Chinese criminal-case judgments show effective charge prediction and relevant-article provision, with reasonable generalization to news-written facts.The news-data result indicates generalization across different expression styles.
2 Related Work
Prior work addressed charge prediction, article identification, case-outcome prediction, legal question answering, and document classification separately or with limited scalability. This paper distinguishes itself by jointly using dynamically guided attention for facts and articles in a multi-label charge task.
- Charge prediction: Earlier charge-prediction studies formulated the task as multi-class classification using KNN or manually designed charge-specific factors.These approaches used shallow textual features or expert knowledge and were difficult to scale across charge types.
- Law-article identification: Prior law-article identification methods either classified fixed article combinations or used a scalable two-step approach.Fixed combinations become difficult to scale as the number of possible article combinations grows exponentially.
- Case outcomes: Case-outcome prediction work targeted binary outcomes such as which party prevails or whether a lower-court decision is affirmed or reversed.
- Case outcomes: This work instead predicts detailed charges, potentially with multiple labels, rather than a binary case outcome.
- Legal question answering: Legal question answering also treats relevant law articles as important, but extracts Japanese Civil Code articles before answering yes/no questions.
- Legal retrieval: Unlike common-law case retrieval, this work predicts specific charges and emphasizes statutory law articles in civil-law decision making.
- Document classification: Compared with standard document classification, the framework also identifies applicable law articles to support and improve charge prediction.
- Method distinction: Its two-stack attention dynamically generates article embeddings from fact-side clues and uses label distributions with validation-tuned thresholds for multi-label prediction.This differs from optimizing separate binary classifiers for each label.
3 Data Preparation
The dataset consists of automatically processed Chinese judgment documents, using document structure and extraction rules to obtain facts, law articles, and charges. The study restricts cases to one defendant to make large-scale construction feasible, leaving multi-defendant matching for future work.
- Data source and split: The corpus was collected from China Judgements Online and split into 50,000 training, 5,000 validation, and 5,000 testing documents.
- Label selection: Only charges appearing more than 80 times in training were classified, while documents with other charges were treated as negative data.
- Document processing: Document indicator clauses divide each judgment into facts, court view, and decision pieces for extracting the corresponding information.
- Document processing: Charges were identified with a manually collected charge list, and law articles were extracted using regular expressions.
- Preprocessing: Charge names were masked from fact descriptions to prevent occasional charge-name leakage into the input.
- Scope boundary: The dataset includes only one-defendant cases because matching facts to charges across defendants is hard and expensive.This simplification may alter the real-world charge distribution but enables an automatically built large-scale, high-quality dataset without legal-practitioner annotation.
4 Our Approach
The approach jointly encodes case facts and relevant law articles, using attention to identify informative content and supportive articles before predicting charges. It trains charge prediction and article attention together, while exposing extracted articles as legal basis.
- Model overview: The model encodes the fact description and concurrently extracts the top k relevant law articles.The extracted articles are embedded separately and passed to an article aggregator.
- Document encoder: Document encoding uses sentence-level and document-level sequence encoders to represent words, sentences, and the whole fact description.The two encoders are implemented with Bi-GRUs in this approach.
- Attentive sequence encoder: Attentive sequence encoding addresses long sequences by weighting informative elements instead of using only endpoint states or treating all elements equally.A context vector generates attention values for sequence elements.
- Article extraction: Article extraction is formulated as 321 binary classifications using word-based SVMs with bag-of-words TF-IDF features, chi-square selection, and linear kernels.Additional articles can be handled by adding corresponding binary classifiers.
- Article encoding and aggregation: The article encoder dynamically generates word- and sentence-level context vectors from each case's fact embedding to focus on case-relevant content.The article aggregator then uses fact-guided attention to select supportive articles and produce an aggregated article embedding.
- Charge prediction and training: The final predictor concatenates fact and article embeddings, applies two fully connected layers and a softmax, and uses a validation-set threshold for positive charges.Training uses cross-entropy for charge prediction and can add a weighted cross-entropy loss supervising article attention with gold-standard articles.
5 Experiments
Experiments compare fact-only and article-aware models for charge prediction, evaluate article extraction, and test robustness on news written in non-legal language. Results show that attention, article supervision, and joint use of facts and law articles improve performance, while noisy extraction and expression differences remain important constraints.
- Experimental Setup: Charge prediction is evaluated with micro- and macro-level precision, recall, and F1 because the top five charges comprise more than 60% of cases.Macro metrics average across charges, whereas micro metrics average across predictions.
- Charge Prediction Results: NN fact outperforms SVM fact by about 4% in micro-F1 using only fact descriptions.The authors attribute this to pretrained embeddings, two-level Bi-GRU encoding, and fact-side attention.
- Charge Prediction Results: 90.21% micro-F1 and 80.48% macro-F1 are achieved by the full model using gold law articles as supervision for article attention.Using noisy extracted articles improves NN fact by around 0.4%, while the SVM fact-plus-article model drops by around 5%.
- Charge Prediction Results: Using only extracted law articles performs worse than using facts, indicating that charge prediction needs both fact descriptions and relevant articles.The authors also suggest that decision logic may be implicitly captured through large numbers of fact-charge pairs.
- Charge Prediction Results: Gold articles improve macro-F1 by more than 30% for SVM and by over 8% for NN relative to corresponding noisy-article settings.These upper-bound comparisons show the importance of law articles and the need to handle noise in extracted articles.
- Article Extraction Results: The article extractor reaches 94.21% recall for the top 20 articles, but its test-set micro-F1 is only 61.08%, motivating attention-based noise handling.Table 2 reports that charge prediction peaks at β = 0.1 because extraction accuracy trades off against classification capacity.
- Performance on News Data: On news data, NN fact supv art achieves 79.12% F1 and has an absolute F1 drop about 10%–15% smaller than SVM fact.The results also show progressive gains from NN fact to NN fact art and then NN fact supv art.
6 Conclusion
The paper presents an attention-based neural network that jointly models charge prediction and relevant article extraction, using weighted articles as legal basis. Experiments show effectiveness on Chinese criminal-case judgments and reasonable generalization to non-professional news descriptions, while multi-defendant cases remain unhandled.
- The proposed framework jointly models charge prediction and relevant article extraction, with weighted relevant articles supporting charge predictions as legal basis.
- Experiments on Chinese criminal-case judgment documents demonstrate the model’s effectiveness for both charge prediction and relevant article extraction.
- Model variants indicate that law articles are important for making judicial decisions in the civil law system.
- Experiments on news data show reasonable generalization from judgment documents to fact descriptions written by non-legal professionals.
- The model cannot explicitly handle multi-defendant cases, and its performance remains below the upper-bound improvement achievable from relevant articles.