Source-linked AI summary
Interpretable Charge Predictions for Criminal Cases: Learning to Generate Court Views from Fact Descriptions
Hai Ye, Xin Jiang, Zhunchen Luo, Wenhan Chao
TL;DR
Charge-prediction systems can produce charges without explanations, while similar case facts can make charge-discriminative rationales difficult to generate. The paper proposes generating court-view rationales from facts with a charge-conditioned Seq2Seq model and reports that experiments show the method’s effectiveness.
Problem
Charge-prediction systems do not provide interpretations, and similar fact descriptions make it difficult for Seq2Seq models to generate charge-discriminative rationales.
Source: §1 · pp. 1–2Method
The paper formulates court-view rationale generation as text-to-text NLG and conditions a Seq2Seq model with attention on encoded charge labels.
Source: §1 · p. 2, §4.2 · p. 4Results
Experiments on a real-world dataset show the model’s effectiveness and that exploiting charge labels improves charge discrimination.
Source: §1 · p. 2Takeaways & Limitations
Generated rationales can support predicted charges and contribute to automatic legal-document generation.
Source: §1 · p. 1The model can generate latent fact details that are not clearly expressed in the fact descriptions.
Source: §5.4 · p. 8
Abstract
from arXiv · showhide
In this paper, we propose to study the problem of COURT VIEW GENeration from the fact description in a criminal case. The task aims to improve the interpretability of charge prediction systems and help automatic legal document generation. We formulate this task as a text-to-text natural language generation (NLG) problem. Sequenceto-sequence model has achieved cutting-edge performances in many NLG tasks. However, due to the non-distinctions of fact descriptions, it is hard for Seq2Seq model to generate charge-discriminative court views. In this work, we explore charge labels to tackle this issue. We propose a label-conditioned Seq2Seq model with attention for this problem, to decode court views conditioned on encoded charge labels. Experimental results show the effectiveness of our method.
1 Introduction
Court-view generation aims to produce judicial rationales from case facts, offering explanations for charges and supporting automatic legal-document generation. Because similar fact descriptions can correspond to different charges, the paper uses charge labels in a label-conditioned Seq2Seq model and releases a real-world dataset for the task.
1 Introduction
- Court-view generation is framed as text-to-text NLG, producing judicial rationales that can support predicted charges and automatic legal-document generation.The task focuses on rationales; charges may be decided separately by judges or charge-prediction systems.Source: §1 · p. 1
- Similar fact descriptions can require different charge-discriminative rationales, such as distinguishing intentional from negligent homicide through killing intent.Unlike a summary of stated facts, a rationale may need to include deduced information not explicit in the description.Source: §1 · pp. 1–2
- The paper proposes the court-view-generation task and releases a real-world dataset for it.The dataset is constructed from Chinese criminal cases.Source: §1 · p. 2
- The proposed label-conditioned Seq2Seq model uses charge labels and attention to generate rationales from case facts.The charge label is encoded to help the decoder select charge-related words.Source: §1 · p. 2
- Experiments on a real-world dataset are reported to show the model’s efficiency and the benefit of charge labels for improving charge discrimination.The supplied passage gives no metric values.Source: §1 · p. 2
2 Related Work
Related work covers legal assistant systems, text-based interpretation, and encoder-decoder generation. The paper positions court-view generation as combining these strands to produce rationales from case facts.
2 Related Work
- Earlier legal assistant work includes charge prediction, identifying applicable articles, legal question answering, and searching for relevant cases.The paper describes charge prediction as text classification and notes that earlier systems do not provide interpretations for their charge decisions.Source: §2 · p. 2
- Court views could contribute to automatic legal-document generation by generating documents step by step from facts obtained from professionals or information extraction.This positions the task alongside legal assistant systems that search, classify, or provide legal information.Source: §2 · p. 2
- Textual explanations have been studied for classification, and the paper presents court views as textual explanations for predicted charges.Prior examples include selecting supportive text snippets for text classification and generating visual explanations for image classification.Source: §2 · p. 3
- The model combines an encoder-decoder approach with charge labels and attention to generate charge-discriminative rationales.The paper relates label use to prior work encoding image labels to generate justification texts.Source: §2 · p. 3
3 COURT-VIEW-GEN Problem
COURT-VIEW-GEN maps case fact descriptions to the rationale portion of a court view, conditioned on the case’s charge. The charge can be determined separately and combined with the generated rationale to form the final court view.
3 COURT-VIEW-GEN Problem
- A court view consists of rationales and a charge, with the rationales supporting the charge.The task generates only the rationale part; the charge may be decided by a person or charge-prediction system and combined with the generated text.Source: §3 · p. 3
- A fact description represents identified facts in a case, such as criminal acts or degree of injury.Source: §3 · p. 3
- The task takes a word sequence of case facts and generates a word sequence of court-view rationales conditioned on charge label v.The objective is to find predicted rationales given the fact description and charge label.Source: §3 · p. 3
4 Our Model
The model generates court-view rationales from fact descriptions, using attention and charge-label conditioning to make the generated text charge-discriminative. Training maximizes rationale likelihood given facts and labels, and inference uses beam search.
4.1 Sequence-to-Sequence Model with Attention
- The baseline encodes fact descriptions with a bidirectional LSTM and decodes rationales with an LSTM using global attention over encoder states.Source: §4.1 · pp. 3–4
4.2 Label-conditioned Sequence-to-Sequence Model with Attention
- Charge-label embeddings join decoder states and context to influence word selection and generate charge-discriminative rationale words.Source: §4.2 · p. 4
- Embedding the charge label in decoder hidden states also shapes attention toward charge-related facts to support more accurate fact details.Source: §4.2 · p. 4
4.3 Model Training and Inference
- Training maximizes the log-likelihood of generated court-view rationales given fact descriptions and charge labels; at inference, rationales are generated by beam search with beam size 5.Source: §4.3 · pp. 4–5
5 Experiments
Experiments compare the model with retrieval and generation baselines, then analyze charge-label effects and generated cases. Results favor the full model while exposing difficulty with latent details.
5.1 Data Preparation
- The dataset comes from China Judgements Online, retains the 50 most frequent charges, and focuses on cases with one defendant and one charge.Names, numbers, and dates are replaced with placeholders; fact descriptions longer than 256 tokens are truncated.Source: §5.1 · p. 5
5.2 Experimental Settings
- Evaluation combines BLEU-4 and variant Rouge scores with human fluency and accuracy ratings from three annotators, plus adoptability counts on 100 outputs per method.Both human rating dimensions use a five-point scale.Source: §5.3 · p. 5
5.3 Comparisons with Baselines
- The full charge-conditioned model performs best overall; attention improves results over retrieval, while its rationales are more accurate than retrieval outputs and comparably fluent.Retrieval methods struggle to match fact details, while generation models produce highly natural rationales.Source: §5.3 · p. 6
5.4 Further Analysis
- Exploiting charge labels improved BLEU-4 for nearly all charges and increased charge-discriminative rationales for every charge; for homicide, the system aligned latent killing motives in rationales with charge labels.The rationale assessment counted charge-relevant rationales among 20 randomly selected candidates.Source: §5.4 · pp. 6–7, Figure 3 · p. 7, Figure 4 · p. 7
- Both charge-label pathways improve performance, and stacking them beats either ablation; removing softmax composition performs better than removing charge composition.The hidden-composition ablation performs similarly to removing charge composition.Source: §5.4 · p. 7, Table 3 · p. 7
5.5 Analysis through Cases
- Case studies show charge labels can preserve input fact details while shifting rationales toward a conditioned charge, but latent theft-frequency details remain difficult.For larceny, the model misses the detail “several times.”Source: §5.5 · p. 8
6 Conclusion and Future Work
The paper introduces court-view rationale generation as text-to-text NLG and uses charge labels in an attention-based Seq2Seq model. Future work includes generating latent details, jointly generating charges and rationales, verification, and more complex cases.
6 Conclusion and Future Work
- The paper formulates court-view generation as text-to-text NLG and proposes an attention-based Seq2Seq model conditioned on charge labels.The authors report that experiments show the effectiveness of their model and of exploiting charge labels.Source: §6 · p. 8
- Future work includes improving latent-detail generation, jointly generating charges and rationales, verifying generated views, and studying cases with multiple charges and defendants.The paper identifies these as future directions.Source: §6 · p. 8
- Figure 8 presents fake-charge-label-conditioned generated rationales in court views and examples of generated rationales.Source: Figure 8 · p. 9