Criminal Court View Generation

• Criminal Court View Generation is a legal natural language generation task that produces the judicial rationale from case materials.
• The task requires preserving factual accuracy and legal consistency by conditioning on variables such as fact descriptions, charge labels, or plaintiff claims.
• Benchmarking efforts include diverse models—from label-conditioned Seq2Seq to event-grounded pipelines—demonstrating practical advances in legal AI.

Searching arXiv for recent and foundational papers on Criminal Court View Generation and closely related court-view generation work. arxiv_search(query="Criminal Court View Generation court view generation legal AI", max_results=10) arxiv_search({"query":"Criminal Court View Generation court view generation legal AI","max_results":10}) Criminal Court View Generation (CVG) is a legal natural language generation task in which a system generates the “Court View”—the court’s reasoning or judicial opinion—from case materials, most commonly a fact description and, in some formulations, additional conditioning variables such as a charge label or plaintiff claims. In the criminal setting, the target text explains why a particular charge is appropriate and how the facts support the legal conclusion; the literature repeatedly distinguishes this output from ordinary summarization because it must preserve factual accuracy, legal consistency, and domain-specific reasoning, while also supporting interpretability and automatic legal document generation (Ye et al., 2018, Hou et al., 16 May 2026, Hou et al., 10 Oct 2025).

1. Conceptual scope and legal function

The core object in CVG is the court’s justificatory text. In the foundational criminal formulation, a court view contains two parts: the rationales, which give the factual and inferential reasons supporting the charge, and the charge itself; the 2018 formulation studies generation of the rationales part only, with the charge assumed to be known in advance (Ye et al., 2018). Later work broadens the task so that a model may generate a full court view from the fact description alone, with the charge then inferred from the generated text or embedded directly in it (Hou et al., 16 May 2026, Hou et al., 10 Oct 2025).

The Chinese legal setting is especially prominent in this literature. One line of work emphasizes that Chinese criminal case documents do not necessarily expose section boundaries through explicit formal headers; instead, the fact section, court view section, and decision section are often identified through recurring lexical cues such as “after identification,” “the court holds that,” and “in accordance with the law” (Hou et al., 10 Oct 2025). This makes CVG a structured legal NLP problem rather than a generic generation problem.

A further complication is that not all papers use exactly the same legal input. Some criminal CVG systems generate from fact descriptions only; some condition on charge labels; and one prominent “Court View Generation” formulation uses both fact descriptions and plaintiff claims, evaluates on a public civil cases dataset, and measures how well the generated text responds to multiple claim types (Li et al., 2024). A common misconception is therefore that all CVG work has a single criminal-only formulation. The published work instead shows a family of closely related tasks centered on judicial explanation generation, with criminal fact-to-rationale generation as the main but not exclusive variant.

2. Task formulations and representational choices

The foundational criminal formulation treats CVG as conditional sequence generation. With fact description $\mathbf{x}$, rationale text $\mathbf{y}$, and charge label $\mathbf{v}$, the task is to generate $\mathbf{y}$ by maximizing $P(\mathbf{y} \mid \mathbf{x}, \mathbf{v})$; the paper explicitly frames this as a text-to-text NLG problem analogous to machine translation or summarization, but with the additional requirement that the output be charge-discriminative (Ye et al., 2018). This distinction is central because criminal fact descriptions are often non-distinct across charges such as intentional homicide and negligent homicide.

A later formulation defines court view generation as: given fact description $f$ and plaintiff claims $c$, generate court view $v$ (Li et al., 2024). In that setting, the court view is judged not only by lexical similarity but also by whether it correctly addresses claim labels such as principal repayment, interest, spousal joint debt, and guarantee liability. The formulation shifts the emphasis from charge discrimination to claim responsiveness, but it preserves the underlying idea that the generated text must express legally salient reasoning rather than surface-level paraphrase.

Event-grounded CVG introduces a more structured intermediate representation. The Event Grounded Generation (EGG) framework defines the task as: given case fact $x$, first extract several events $e$ from the fact, then generate the court-view rationales $\mathbf{y}$0, while gold events are unavailable. The extracted events are concatenated with the fact as $\mathbf{y}$1, so generation is grounded in fine-grained legal events rather than only coarse fact segments (Yue et al., 2024). This reflects a representational shift from document-level conditioning to event-level conditioning.

Recent lightweight-LLM work adds another representational contrast: “CVG First” versus “Directly.” In the CVG-first setting, the model is prompted to generate a court view and then predict the charge from that reasoning, using the output format <view> court view </view>, <charge> charge </charge>. In the direct setting, it outputs only <charge>charge</charge> (Hou et al., 16 May 2026). Here the court view becomes an explicit intermediate reasoning artifact rather than only a terminal generation target.

3. Datasets and benchmark infrastructures

The early criminal CVG corpus is built from China Judgements Online and restricts attention to cases with one defendant and one charge. Its reported statistics are 153,706 training cases, 9,152 development cases, and 9,123 test cases, with average fact length 219.9 tokens, average rationale length 30.6 tokens, and 51 charge labels after keeping the top 50 charges by frequency and mapping others to “others” (Ye et al., 2018). This dataset established the standard fact-to-rationale benchmark for charge-conditioned generation.

Subsequent work expanded both scale and task structure. The CCVG dataset is collected from China Judgments Online, covers cases from 1985 to 2021, and contains more than 110K cases after extracting fact and court-view sections, extracting charges, filtering out samples shorter than 50 characters or longer than 512 characters, and splitting the data at a 9:1 ratio (Hou et al., 10 Oct 2025). The same paper reports that 0.2% of training examples have unsuccessfully extracted charge labels and are kept with the charge field set to null, while all test samples have non-empty fields.

CVGEvalKit consolidates three public datasets into a mixed-training and cross-dataset evaluation regime: C3VG, LCVG, and CCVG. The reported splits are 50,312 training and 12,627 test instances for C3VG; 60,744 training and 20,290 test instances for LCVG; and 111,252 training and 1,154 test instances for CCVG (Hou et al., 16 May 2026). This benchmark infrastructure is important because it explicitly studies cross-dataset generalization rather than only single-dataset fitting.

Event-grounded CVG uses a different resource configuration. Its court-view generation dataset, also derived from China Judgments Online, contains 62,939 samples across 62 case types, with average fact length 458.1, average court view length 130.9, average annotated event questions per case type 9.0, and average total event length per case 83.4. For event extraction, it uses CJRC, a Chinese judicial reading comprehension dataset with 20,000 samples, average paragraph length 501.8, and average question length 16.5 (Yue et al., 2024).

The claim-responsive line uses a public civil cases dataset adapted from Wu et al. (2020), with average fact description length 177.5 tokens, average plaintiff claims length 76.3 tokens, average court view length 205.7 tokens, average number of claim labels 2.13, minimum lengths of 20 tokens for fact, claim, and court view, maximum lengths 400 / 200 / 400, and an 8:1:1 train/validation/test split (Li et al., 2024). Together, these datasets show that “court view generation” is benchmarked under materially different legal-document regimes.

4. Model families and methodological evolution

The first arXiv CVG model is a label-conditioned Seq2Seq model with attention. It uses a bidirectional LSTM encoder, an LSTM decoder, and global attention following Luong et al.; charge information is injected both into the decoder hidden state and into the output distribution through a learned charge-label embedding (Ye et al., 2018). The stated rationale is that charge labels restrict the generation space, bias the decoder toward charge-related words and legal reasoning, and help the system generate more charge-discriminative rationales.

Knowledge-guided prompt methods replace full sequence-model retraining with parameter-efficient conditioning. The Knowledge Injection and Guidance (KIG) framework keeps the PLM frozen and trains only a knowledge-injected prompt encoder. Its two knowledge sources are keyword knowledge for each claim label and label-definition knowledge from the Code of Law. At inference time, a generating navigator—implemented as a multi-label claim classifier—adjusts token probabilities so that the partial generation remains aligned with the correct claim labels (Li et al., 2024). This design separates knowledge acquisition during training from knowledge usage during decoding.

Event-grounded models introduce an extract-then-generate pipeline. EGG first uses an LLM-based QA-style event extractor to answer case-type-specific event questions, then concatenates the resulting event descriptions with the case fact and feeds the combined input to BART for rationale generation. The companion variant, $\mathbf{y}$2, still uses LLM-based extraction during training but removes the need for event extraction during inference by treating event information as auxiliary supervision and aligning fact and event representations with a contrastive objective (Yue et al., 2024). The central methodological claim is that fine-grained event grounding is more effective than coarse fact segmentation.

The lightweight-LLM line reframes CVG as a testbed for sub-2B instruction-following models. One paper compares Qwen2.5-0.5B-Instruct, Qwen3-0.6B, Qwen3-1.7B, Llama-3.2-1B-Instruct, Gemma3-1B-IT, and InternLM2.5-1.8B-Chat against DNN baselines such as AttS2S, Transformer, and C3VG, and against prior legal LLMs such as LeGen-PT, LeGen-FT, and ShiZhi. Fine-tuning uses LoRA with rank $\mathbf{y}$3, $\mathbf{y}$4, and dropout $\mathbf{y}$5 (Hou et al., 16 May 2026). This line treats CVG both as a generation task and as an intermediate representation for charge prediction.

A specialized domain model is ShiZhi, described as the first LLM specifically designed for court view generation. It is a fine-tuned version of Qwen2-0.5B-Instruct, trained for 3 epochs with batch size 4, learning rate $\mathbf{y}$6, maximum input length 512 tokens, on a single NVIDIA RTX 4090 GPU, using the Swift framework (Hou et al., 10 Oct 2025). The training format uses a legal-assistant-style prompt in which the system is instructed to assume the role of a judge and generate the corresponding court view section from the fact section.

5. Evaluation protocols and empirical findings

Evaluation in CVG is multi-layered. The standard generation metrics are ROUGE-1, ROUGE-2, ROUGE-L, and BLEU variants; charge prediction is usually measured with Accuracy and Macro-F1; claim-responsive work adds Micro-F1, Macro-F1, Micro-Jaccard, and Macro-Jaccard computed by feeding generated court views into a trained claim-label classifier; event-grounded work also reports BERTScore and Macro-average F1 for event extraction (Ye et al., 2018, Li et al., 2024, Yue et al., 2024, Hou et al., 16 May 2026).

The 2018 label-conditioned Seq2Seq model established that explicit charge conditioning improves both automatic metrics and human judgments. Its reported automatic scores are BLEU-4 45.8, ROUGE-1 70.9, ROUGE-2 52.5, and ROUGE-L 67.7, outperforming retrieval baselines, phrase-based statistical MT, standard Seq2Seq, and an attention-based abstractive summarization baseline. In human evaluation it achieves fluency 4.93, accuracy 4.54, and adoptability 0.72 (Ye et al., 2018). The paper’s ablations show that removing charge conditioning from either the hidden-state component or the output-layer component degrades performance.

KIG reports the best performance on all reported metrics in its setting. The headline numbers are BLEU-1 71.04, BLEU-2 64.80, BLEU-N 62.40, ROUGE-L 71.11, Micro-F1 90.21, Macro-F1 87.78, Micro-Jaccard 82.16, and Macro-Jaccard 78.40, with a stated 11.87% improvement in Macro-F1 over the best baseline. In human evaluation over 200 test cases with 10 legal-background annotators, KIG obtains Fluency 4.52 and Fitness 4.02 (Li et al., 2024). Its ablations indicate that keyword-based initialization, label attention, and the generating navigator each contribute.

Event grounding improves over coarse segmentation. On the reported benchmark, EGG achieves ROUGE-1 76.86, ROUGE-2 65.15, ROUGE-L 65.90, BLEU-1 56.92, BLEU-2 54.43, BLEU-N 53.59, and BERTScore 86.54; $\mathbf{y}$7 reaches ROUGE-1 75.23, ROUGE-2 64.41, ROUGE-L 64.24, BLEU-1 57.19, BLEU-2 54.34, BLEU-N 53.41, and BERTScore 85.86; both outperform C3VG with BART in most metrics. The event extractor obtains Macro-F1 84.6 on CJRC test, higher than the original CJRC BERT-based result of 82.9. The efficiency table reports Baichuan-7B(Fact) at 1.00×, $\mathbf{y}$8 at 0.71×, and EGG at 8.72×, and the paper interprets this as $\mathbf{y}$9 being roughly 12× faster than EGG during decoding/inference (Yue et al., 2024).

ShiZhi demonstrates the effect of domain-specific fine-tuning on a small model. Against Qwen2-0.5B-Instruct, it raises BLEU-1 from 11.1 to 58.5, BLEU-2 from 6.3 to 51.0, and BLEU-N from 2.9 to 41.7; ROUGE-1, ROUGE-2, and ROUGE-L rise from 0.0, 0.0, and 0.0 to 6.3, 0.8, and 6.3. For charge prediction, the baseline reports Accuracy 17.3 and Macro F1 29.5, while ShiZhi is reported at about 86% Accuracy and about 92.5–92.8 Macro F1; the paper explicitly notes a discrepancy between the table, which gives 86.5 and 92.8, and the abstract and narrative, which give 86.1% and 92.5% (Hou et al., 10 Oct 2025).

The lightweight-LLM benchmark adds several important negative results. Architecture matters substantially: among zero-shot models, Qwen3-1.7B generally outperforms Llama-3.2-1B-Instruct and InternLM2.5-1.8B-Chat, while Gemma3-1B-IT performs relatively well in CVG but often produces charges that are semantically close without using the exact legal terminology expected by the benchmark. Size helps, but not uniformly; after fine-tuning, the performance gap between Qwen3-0.6B and Qwen3-1.7B narrows, and the authors explicitly avoid architecture-agnostic scaling claims below 2B. Fine-tuned lightweight LLMs outperform DNN baselines: on C3VG, InternLM2.5-1.8B-Chat(FT) achieves 78.68 ROUGE-L and 67.31 BLEU-n, described as roughly 32% and 66% gains over the best DNN baseline, C3VG, and exceeds it by about 3% Accuracy and 36% Macro-F1 in charge prediction. Most notably, CVG First is not universally beneficial: it helps smaller models such as Qwen3-0.6B and Gemma3-1B-IT most clearly, but can hurt larger lightweight models such as Qwen3-1.7B and InternLM2.5-1.8B-Chat (Hou et al., 16 May 2026). This directly contradicts the simplistic view that explicit intermediate reasoning always improves legal classification.

6. Provenance, structured case views, and open issues

Narrative court-view generation is only one way of representing legal reasoning. Visilant, although not itself a criminal court view generation system, provides a closely related investigation interface organized around “visualization documents,” analysis states, branches, merges, and a report viewer. It supports a workflow from case overview to exploration, hypothesis generation, comparison and merge, and report generation; it also distinguishes three credibility levels—Evidence, Knowledge, and Assumptions—and visually marks states affected by updates to underlying data (Zákopčanová et al., 2020). In a court-facing context, its progress tracking diagram can reconstruct how a conclusion was formed, and its report-generation mode can assemble a presentation-ready report for state attorneys. This suggests a provenance-aware front end for CVG, although that implication is architectural rather than explicitly claimed in the paper.

A second adjacent line converts judicial prose into structured graphs rather than generating narrative rationales. The French cassation study introduces a domain-specific criminal ontology for appeals to the Cour de cassation and an ontology-guided LLM pipeline that outputs Turtle-formatted RDF triples, validates them with rdflib, and stores them in Fuseki. On a 20-appeal expert-validated ground-truth set, GPT-4o mini reaches precision 93% and recall 89%; on the full 2023 dataset, 2820 appeals are processed with an average of 30 triples per appeal (Belikov et al., 24 Jan 2025). The resulting “view” preserves participants, offense types, punishments, appeal decision types, legal articles, and procedural links. This is not CVG in the narrative sense, but it demonstrates that criminal court “views” can also be formalized as structured legal abstractions.

The main limitations reported across the literature are jurisdictional, temporal, and representational. ShiZhi and CCVG are Chinese-only, CCVG includes cases only up to 2021, and only the 0.5B scale is studied (Hou et al., 10 Oct 2025). KIG notes knowledge freshness, the need for extra training and labeled data for the navigator, domain and jurisdiction specificity, and bias and privacy concerns (Li et al., 2024). EGG notes that events are currently obtained by simple post-processing of question-answer pairs and that complex inter-event relations such as causality are not modeled well, motivating future event graphs (Yue et al., 2024). The French KG pipeline reports malformed RDF, typed-literal errors, ontology projection errors, and a trade-off between ontology size and prompt tractability (Belikov et al., 24 Jan 2025). The lightweight-LLM benchmark adds a broader caution: these systems are suitable for research or supportive purposes, not as authoritative judges (Hou et al., 16 May 2026).

Taken together, the literature defines CVG as a family of judicial-explanation generation tasks spanning charge-conditioned rationale generation, fact-to-court-view generation, claim-responsive court-view generation, and event-grounded legal generation. A plausible implication is that future systems will combine three strands that are currently separated: narrative generation for explanation, structured extraction for normalization and retrieval, and provenance tracking for courtroom traceability.