AppealCase: A Dataset and Benchmark for Civil Case Appeal Scenarios

Ang Li; Changlong Sun; Fei Wu; Keting Yin; Kun Kuang; Meitong Guo; Xiaozhong Liu; Yiquan Wu; Yuting Huang

REVIEW 3 major objections 3 minor 45 references

AppealCase: A Dataset and Benchmark for Civil Case Appeal Scenarios

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Appeal reversal prediction stumps all 20 LLMs in new test

desk verdict A genuinely useful new resource for appellate LegalAI, but the second-instance benchmark leaks the answer and the abstract overstates the results; fixable, but not as is. read the letter →

arxiv 2505.16514 v2 pith:DYOMDYCJ submitted 2025-05-22 cs.CL

Yuting Huang , Meitong Guo , Yiquan Wu , Ang Li , Xiaozhong Liu , Keting Yin , Changlong Sun , Fei Wu

show 1 more author

Kun Kuang

This is my paper

classification cs.CL

keywords AppealCaselegaljudgmentpredictionappellatereviewcivilproceduredatasetLLMbenchmarkChinesecasesreversal

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces AppealCase, a dataset of 10,000 matched pairs of real Chinese civil first-instance and second-instance judgments, annotated along five dimensions: whether the judgment was reversed, the reversal reasons, cited legal provisions, claim-level decisions, and whether new information appeared in the second instance. On this foundation it defines five LegalAI tasks and evaluates 20 large language models. The paper's central empirical claim is that no current model reaches 50% macro-F1 on predicting reversal reasons from the first-instance perspective, and only a few exceed 50% from the second-instance perspective, so appellate review remains a largely unsolved problem for legal artificial intelligence. A sympathetic reading takes the contribution as the resource itself plus the demonstration of a persistent performance gap.

What carries the argument

The central object is the matched first-instance and second-instance judgment pair, structured into header, claim, fact description, court's view, judgment, and footer, with five annotation layers. The task design splits judgment reversal prediction into a first-instance perspective (first-instance document plus second-instance claim) and a second-instance perspective (additionally including the second-instance fact description with any new information). The annotation pipeline combines rules, repeated LLM sampling, and human expert review, with three legal experts checking 500 cases. This machinery turns unstructured court documents into a testbed for appellate reasoning.

What would settle it

Select a random sample of the 4,519 second-instance-perspective cases, strip or mask any clause in the fact description that states whether the first-instance judgment was wrong or reversed, and re-run the same models; if F1 falls toward the first-instance-perspective scores, the second-instance benchmark is inflated by leakage.

Watch

Extended reading notes

Core claim

The paper's central discovery is that current models cannot yet predict whether and why an appellate court overturns a first-instance civil judgment. Across 20 non-reasoning, reasoning, and legal-domain large language models, macro-F1 on the judgment reversal prediction task stays below 50% in the first-instance perspective; in the second-instance perspective, more than half of the models also fall below 50%, with the best score at 57.40%. The paper attributes the difficulty to long documents, dynamic legal standards, easily confused legal relationships, and nuanced liability allocations, and it reports that fine-tuning a small model improves scores but keeps them below 60%.

Load-bearing premise

The load-bearing premise is that the second-instance perspective input contains facts but not the second-instance outcome; the paper's own Appendix B.2 example already states that the first-instance finding is contrary to objective facts and legal provisions, so if that phrasing is typical, the second-instance results partly reflect label leakage rather than appellate reasoning.

Editorial extensions

If this is right

Reversal prediction could give first-instance courts a warning signal and help reduce misjudgment rates.
Legal provision recommendation and claim-level judgment prediction could assist second-instance courts in reviewing appeals.
Court view generation could help second-instance courts draft judgment documents by providing the appellate reasoning structure.
The benchmark results imply that current large language models are not yet reliable enough for appellate assistance in civil cases.
The dataset's 91 civil causes of action create a reusable testbed for studying how facts, claims, and court views evolve across trial levels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the five-dimension annotation schema is tied to civil procedure, it could be adapted to criminal and administrative appeals to test whether the observed model failures are specific to civil cases.
The hard-case analysis points to explicit modeling of legal relationship classification (for example, loan versus partnership) as a promising decomposition before reversal prediction; the dataset's claim-level annotations make that decomposition testable.
The paired-document structure could support a second-instance rationale generation evaluation that conditions only on the first-instance document, isolating the model's ability to transfer legal reasoning across instances.
Since most universally missed reversal cases involve legal application errors, a focused benchmark subset isolating that category would sharpen the diagnostic value of the dataset.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The AppealCase dataset is the first thing I've seen that pairs first- and second-instance Chinese civil judgments at scale (10k pairs, 91 causes of action), with annotations for reversals, reasons, legal provisions, claim-level outcomes, and new evidence. That is a real gap: every prior LegalAI benchmark I know is single-stage. The first-instance reversal prediction task, where the model reads only the trial record and the appeal request, is a genuinely new setup and does not obviously leak the answer. The dataset is slated for release under CC BY-NC, and the authors report a human-expert check on 500 cases with >99% agreement, which is encouraging if the dataset is actually released.

But the paper as it stands has a load-bearing flaw. Problem 2 (second-instance perspective) gives the model the second-instance fact description as input. In real Chinese judgments, that section often states the appellate court's own finding—including whether the first-instance finding was wrong. Appendix B.2 is explicit: the input says 'the first-instance court's finding... is contrary to objective facts and legal provisions,' and the gold label is exactly 'Factual determination errors, Legal application errors.' So the model can copy the answer. This is not a minor leakage; it makes the second-instance benchmark circular.

There are also two numerical inconsistencies that need to be cleared up. Section 2.1 says 50% of the 10,000 cases were reversed but only 240 have reasons for reversal. Those numbers cannot both be right if reversal reason is the label for the headline task. Table 3 reports 5,481 and 4,519 samples for the two perspectives, which sum to 10,000 but do not tell us how many reversed cases carry reasons. And the abstract's claim that 'all current models achieve less than 50% F1' is contradicted by Table 4, where several models exceed 50% on the second-instance perspective and fine-tuned Qwen3 reaches 63.46. The text in Section 4.2 gets this right, so the abstract is just sloppy, but it is still the headline.

Minor issues: the fine-tuning section calls the model Qwen3-0.6b and later Qwen3-0.8B; no variance is reported. None of these matter compared to the leakage.

My take: the dataset is worth a serious referee, not a desk reject. The resource is novel and the first-instance task is usable. But I would not use the second-instance benchmark until the authors re-define the inputs to exclude the appellate court's conclusion, fix the reversal-count contradiction, and correct the abstract. If they do that, I'd happily cite it.

Referee Report

3 major / 3 minor

Summary. AppealCase is a proposed dataset of 10,000 matched first-instance and second-instance civil judgment document pairs obtained from China Judgments Online, spanning 91 causes of action. The authors annotate each case on five dimensions: judgment reversal, reversal reasons, cited legal provisions, claim-level support decisions, and presence of new information. Based on these annotations, they define five supervised tasks: two variants of judgment reversal prediction (first-instance and second-instance perspectives), legal provision recommendation, legal judgment prediction, and court view generation. The paper reports evaluations of 20 large language models plus two fine-tuned small models, and claims that all current models achieve under 50% F1 on judgment reversal prediction, which it interprets as evidence that appellate scenario modeling remains unsolved.

Significance. If the benchmark is sound, AppealCase would be a useful new resource for LegalAI, filling a gap for appellate tasks beyond first-instance judgment prediction. Its grounding in public court records, the matched-pair structure, expert evaluation of a 500-case sample with reported 99% accuracy, and the breadth of the model evaluation are credit-worthy strengths. The five tasks are well-motivated and the dataset could support research on judicial consistency. However, the validity of the headline result and of one of the two reversal-prediction settings depends on resolving the label-leakage, annotation-count, and abstract-accuracy issues raised below. The other three tasks (provision recommendation, judgment prediction, court view generation) are less affected by these concerns.

major comments (3)

[Section 3.1 (Problem 2); Appendix B.2] The second-instance perspective task as defined has label leakage. The input includes the second-instance fact description, and the example in Appendix B.2 contains the sentence "The first-instance court's finding ... is contrary to objective facts and legal provisions," which directly states the target label (factual determination errors and legal application errors). A model can copy this conclusion from the input to produce the output ("Factual determination errors, Legal application errors") without performing any appellate reasoning. If this pattern is systematic in the dataset, the F1 scores in Table 4 for this perspective (e.g., 57.40 for Qwen2.5-72B) overestimate true predictive ability and the benchmark is circular. The task should be redefined to provide only the new evidence or factual additions from the second instance, with the appellate court's evaluative statements and conclusions removed from the input. A similar concern applies to Problem 4 (Section 3.3): the second-instance fact description in Appendix B.4 contains "This court finds as follows..." statements that may leak the claim-support label.
[Section 2.1] The sentence "Among these 10,000 cases, 50% were reversed, and there are 240 cases with reasons for reversal" is internally contradictory: 50% of 10,000 is 5,000, not 240. Table 3 lists 5,481 and 4,519 samples for the two judgment-reversal-prediction perspectives, implying that reversal reasons are available for the large majority of the 10,000 cases. As written, the manuscript cannot be used to reconstruct the benchmark because it is unclear how many reversed cases actually carry the "Reasons for Reversal" label. This must be clarified, and the dataset statistics should be made consistent between the prose, Table 2, and Table 3.
[Abstract; Section 4.2; Table 4] The abstract claims that "all current models achieve less than 50% F1 scores on the judgment reversal prediction task," but Table 4 shows that nine of the twenty models exceed 50% F1 on the second-instance perspective, e.g., Qwen2.5-72B at 57.40, DeepSeek-V3 at 54.49, Doubao-1.5-pro at 54.73, and DeepSeek-R1 at 54.77. Section 4.2 itself correctly notes that "more than half of the models have F1 scores below 50%" for that perspective, contradicting the abstract. The headline claim should be restricted to the first-instance perspective or to an average across perspectives, and the wording must be aligned with the reported numbers.

minor comments (3)

[Section C.2 / Section 4.2] The fine-tuned model is called "Qwen3-0.6b" in Section 4.2 but "Qwen3-0.8B" in Section C.2; the model name and size should be consistent, and the release name should be verified.
[Section 2.2.2, Judgment Reversal] The rule for labeling judgment reversal uses the keywords "appeal dismissed" and "original judgment upheld" to mark non-reversed cases and "otherwise" to mark reversed. It is not specified how cases are handled when only one of the two keywords appears, or when the judgment is partially reversed; additional details on the rule and its failure cases would improve reproducibility.
[Section 1] The phrase "is largely remains unexplored" is ungrammatical; it should read "remains largely unexplored" or similar. Also, the caption of Table 3 states "Type # Sample" with an inconsistent use of the "#" character; consider using "Task Type" and "Samples".

Circularity Check

1 steps flagged · score 6.0 of 10

Second-instance reversal-prediction task leaks its target label into the supplied fact description, so that half of the headline benchmark reduces to copying.

self definitional [Section 3.1 (Problem 2) and Appendix B.2]
"Given the first-instance document and the second-instance claim and fact description, which contains new information introduced in the second instance, the task is to predict the reasons for reversal. ... The first-instance court’s finding that the appellant had no evidence to prove that he was injured while working on Building 3, Section 1 of the Banlishan Resettlement Housing Project is contrary to objective facts and legal provisions. ... Output: Factual determination errors, Legal application errors"

The task input includes the second-instance fact description taken from the final appellate judgment. In the supplied example, that fact description states that the first-instance finding is 'contrary to objective facts and legal provisions,' which is precisely the target label 'Factual determination errors, Legal application errors.' Thus the model is not predicting an unseen reversal outcome; it is copying the appellate court's own conclusion from the input. The reported second-instance F1 scores therefore incorporate an answer-present leakage and cannot support the paper's claim about the difficulty of appellate reversal prediction.

full rationale

The external dataset construction is not circular: reversal labels are extracted from public court documents via rules and expert review (Section 2.2.2), and Problem 1, Court View Generation, Legal Judgment Prediction, and the annotation tasks are grounded in the judgments themselves. The circularity is localized to the design of Problem 2, where the supplied second-instance 'fact description' contains the appellate conclusion that is the target. Appendix B.2 makes the leak explicit. Since this is the paper's headline reversal-prediction benchmark, the construction-level circularity is material: it invalidates that sub-task as an independent test of LegalAI appellate reasoning. Separately, the manuscript contains numerical inconsistencies—Section 2.1's '240 cases with reasons for reversal' versus the 50% reversal rate and the 5,481/4,519 sample counts in Table 3, and the abstract's 'less than 50% F1' versus Table 4's >50% scores. These are correctness issues, not circularity, and should be corrected. Score 6: one headline task reduces by construction, while the underlying corpus and the remaining tasks retain independent content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No physical or theoretical entities are invented. The load-bearing assumptions are about the reliability of rule-based and LLM-based annotation of Chinese court documents, plus the implicit assumption that task inputs do not leak the target label.

free parameters (3)

Reversal sampling ratio = 50%
Dataset composition chosen by the authors, not derived from population appeal rates; affects class balance and baseline F1.
LLM annotation consistency threshold = 10 identical responses required; low consistency sent to manual review
Hand-chosen rule in Section 2.2.2; the boundary between all-identical and low-consistency is not quantified.
Legal provision distractor selection = random selection from other cases
Distractors in the recommendation task are randomly chosen without a stated seed or distance metric, affecting task difficulty.

assumptions (4)

domain assumption Chinese judgment documents follow six predictable sections and keyword patterns that reliably segment claims and reversal language.
Used in Section 2.2.1 for document structuring and in the keyword-based reversal label rules.
domain assumption A second-instance judgment is 'not reversed' iff its judgment paragraph contains both keywords 'appeal dismissed' and 'original judgment upheld'.
Rule in Section 2.2.2; partial reversals or differently phrased outcomes would be mislabeled.
domain assumption LLM annotations that are identical across ten sampling rounds are treated as correct without expert verification for all 10,000 cases.
Section 2.2.2; only low-consistency samples go to manual review.
domain assumption The 500-case expert review validates the remaining 9,500 cases by extension.
Section 2.3 reports over 99% accuracy on the sample, assumed to generalize to the full dataset.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AppealCase: A Dataset and Benchmark for Civil Case Appeal Scenarios." pith.science (2026). https://pith.science/paper/DYOMDYCJ

@misc{pith2026250516514,
  author       = {Pith},
  title        = {Pith review of: AppealCase: A Dataset and Benchmark for Civil Case Appeal Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DYOMDYCJ}},
  note         = {Machine review of arXiv:2505.16514}
}

read the original abstract

Recent advances in LegalAI have primarily focused on individual case judgment analysis, often overlooking the critical appellate process within the judicial system. Appeals serve as a core mechanism for error correction and ensuring fair trials, making them highly significant both in practice and in research. To address this gap, we present the AppealCase dataset, consisting of 10,000 pairs of real-world, matched first-instance and second-instance documents across 91 categories of civil cases. The dataset also includes detailed annotations along five dimensions central to appellate review: judgment reversals, reversal reasons, cited legal provisions, claim-level decisions, and whether there is new information in the second instance. Based on these annotations, we propose five novel LegalAI tasks and conduct a comprehensive evaluation across 20 mainstream models. Experimental results reveal that all current models achieve less than 50% F1 scores on the judgment reversal prediction task, highlighting the complexity and challenge of the appeal scenario. We hope that the AppealCase dataset will spur further research in LegalAI for appellate case analysis and contribute to improving consistency in judicial decision-making.

Figures

Figures reproduced from arXiv: 2505.16514 by the authors.

**Figure 2.** A comparative example of first-instance and second-instance documents. The diagram illustrates the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Distribution of reversal reasons in the Appeal [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Performance of different models on the five tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 25 canonical work pages

[1]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Bodenheimer, Zhenlai Deng, and Jingwu Ji

E. Bodenheimer, Zhenlai Deng, and Jingwu Ji. 1987. Jurisprudence: Philosophy of Law and Its Methods. Huaxia Publishing House

work page 1987
[4]

Marius B \"u ttner and Ivan Habernal. 2024. https://aclanthology.org/2024.eacl-long.122/ Answering legal questions from laymen in G erman civil law system . In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2015--2027, St. Julian ' s, Malta. Association for Computa...

2024
[5]

Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, and Ion Androutsopoulos. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.261 LEGAL - BERT : The muppets straight out of law school . In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2898--2904, Online. Association for Computational Linguistics

doi:10.18653/v1/2020.findings-emnlp.261 2020
[6]

Junyun Cui, Xiaoyu Shen, Feiping Nie, Zheng Wang, Jinglong Wang, and Yulong Chen. 2022. https://doi.org/10.48550/arXiv.2204.04859 A survey on legal judgment prediction: Datasets, metrics, models and challenges . CoRR, abs/2204.04859

arXiv doi:10.48550/arxiv.2204.04859 2022
[7]

Junyun Cui, Xiaoyu Shen, and Shaochun Wen. 2023. A survey on legal judgment prediction: Datasets, metrics, models and challenges. IEEE Access, 11:102050--102071

2023
[8]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long a...

doi:10.18653/v1/n19-1423 2019

Show all 45 references

[9]

Zhiwei Fei, Xiaoyu Shen, Dawei Zhu, Fengzhe Zhou, Zhuo Han, Alan Huang, Songyang Zhang, Kai Chen, Zhixin Yin, Zongwen Shen, et al. 2024. Lawbench: Benchmarking legal knowledge of large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Langu...

2024
[10]

Yi Feng, Chuanyi Li, and Vincent Ng. 2022. Legal judgment prediction: A survey of the state of the art. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22, pages 5461--5469

2022
[11]

Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793

2024 arXiv
[12]

Neel Guha, Julian Nyarko, Daniel Ho, Christopher R \'e , Adam Chilton, Alex Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel Rockmore, Diego Zambrano, et al. 2023. Legalbench: A collaboratively built benchmark for measuring legal reasoning in large language models. Advances...

2023
[13]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

2025 arXiv
[14]

Wei Jiang and Ming Shao. 1998. Achievements and issues in the study of civil procedure law and its development. China Law Review, (4):18--22

1998
[15]

Jeffrey R Lax. 2007. Constructing legal rules on appellate courts. American Political Science Review, 101(3):591--604

2007
[16]

Ang Li, Qiangchao Chen, Yiquan Wu, Ming Cai, Xiang Zhou, Fei Wu, and Kun Kuang. 2024 a . http://arxiv.org/abs/2403.04369 From graph to word bag: Introducing domain knowledge to confusing charge prediction

2024 arXiv
[17]

Ang Li, Yiquan Wu, Yifei Liu, Kun Kuang, Fei Wu, and Ming Cai. 2024 b . https://aclanthology.org/2024.lrec-main.522/ Enhancing court view generation with knowledge injection and guidance . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, ...

2024
[18]

Haitao Li, Junjie Chen, Jingli Yang, Qingyao Ai, Wei Jia, Youfeng Liu, Kai Lin, Yueyue Wu, Guozhi Yuan, Yiran Hu, et al. 2024 c . Legalagentbench: Evaluating llm agents in legal domain. arXiv preprint arXiv:2412.17259

2024 arXiv
[19]

Haitao Li, You Chen, Qingyao Ai, Yueyue Wu, Ruizhe Zhang, and Yiqun Liu. 2024 d . Lexeval: A comprehensive chinese legal benchmark for evaluating large language models. In Proceedings of the Thirty-eighth Conference on Neural Information Processing Systems (NeurIPS 2024), Data...

2024
[20]

Hao Li. 2008. The allocation of burden of proof in civil judgments: An analysis based on the *gazette* case sample. Tsinghua Law Review, 6

2008
[21]

Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

2024 arXiv
[22]

John Merryman and Rogelio P \'e rez-Perdomo. 2018. The civil law tradition: an introduction to the legal systems of Europe and Latin America. Stanford University Press

2018
[23]

Ni Ming. 2021. A study on the second-instance ruling in china's civil procedure. Master's thesis, Soochow University

2021
[24]

Weicong Qin, Zelin Cao, Weijie Yu, Zihua Si, Sirui Chen, and Jun Xu. 2024. Explicitly integrating judgment prediction with legal document retrieval: A law-guided generative approach. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in I...

2024
[25]

Tyss Santosh, Shanshan Xu, Oana Ichim, and Matthias Grabmair. 2022. Deconfounding legal judgment prediction for european court of human rights cases towards better alignment with experts. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing...

2022
[26]

Guy I Seidman. 2016. The new comparative civil procedure. In The Dynamism of Civil Procedure-Global Trends and Developments, pages 19--44. Springer

2016
[27]

Sergot, Fariba Sadri, Robert A

Marek J. Sergot, Fariba Sadri, Robert A. Kowalski, Frank Kriwaczek, Peter Hammond, and H Terese Cory. 1986. The british nationality act as a logic program. Communications of the ACM, 29(5):370--386

1986
[28]

Jingyun Sun, Shaobin Huang, and Chi Wei. 2024. Chinese legal judgment prediction via knowledgeable prompt learning. Expert Systems with Applications, 238:122177

2024
[29]

Suxin Tong, Jingling Yuan, Peiliang Zhang, and Lin Li. 2024. Legal judgment prediction via graph boosting with constraints. Information Processing & Management, 61(3):103663

2024
[30]

Nirmalie Wiratunga, Ramitha Abeyratne, Lasal Jayawardena, Kyle Martin, Stewart Massie, Ikechukwu Nkisi-Orji, Ruvan Weerasinghe, Anne Liret, and Bruno Fleisch. 2024. Cbr-rag: case-based reasoning for retrieval augmented generation in llms for legal question answering. In Intern...

2024
[31]

Yiquan Wu, Kun Kuang, Yating Zhang, Xiaozhong Liu, Changlong Sun, Jun Xiao, Yueting Zhuang, Luo Si, and Fei Wu. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.56 De-biased court`s view generation with causality . In Proceedings of the 2020 Conference on Empirical Methods in...

2020 doi
[32]

Yuemin Xi. 2018. The standards and rules for second-instance ruling changes in civil and commercial cases. Journal of China University of Political Science and Law, (3):110--124

2018
[33]

Chaojun Xiao, Xueyu Hu, Zhiyuan Liu, Cunchao Tu, and Maosong Sun. 2021. Lawformer: A pre-trained language model for chinese legal long documents. AI Open, 2:79--84

2021
[34]

Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. 2023. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305

2023 arXiv
[35]

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

2025 arXiv
[36]

an Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115

2024 arXiv
[37]

Shengbin Yue, Wei Chen, Siyuan Wang, Bingxuan Li, Chenchen Shen, Shujun Liu, Yuxuan Zhou, Yao Xiao, Song Yun, Xuanjing Huang, et al. 2023. Disc-lawllm: Fine-tuning large language models for intelligent legal services. arXiv preprint arXiv:2309.11325

2023 arXiv
[38]

Weiping Zhang. 2012. The value trade-off between efficiency and fairness in the revision of the *civil procedure law*. Chinese Judiciary, (6):29--34

2012
[39]

Zhichao Zhao. 2025. The authoritative logic behind the explanation system: Defending active clarification. Journal of China University of Political Science and Law, (02):148--159

2025
[40]

Min Zheng, Bo Liu, and Le Sun. 2022. Lawrec: automatic recommendation of legal provisions based on legal text analysis. Computational Intelligence and Neuroscience, 2022(1):6313161

2022
[41]

Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. 2024. https://doi.org/10.18653/v1/2024.acl-demos.38 L lama F actory: Unified efficient fine-tuning of 100+ language models . In Proceedings of the 62nd Annual Meeting of the Association for Computational Lin...

2024 doi
[42]

ZhihaiLLM. 2023. https://huggingface.co/ZhihaiLLM/wisdomInterrogatory W isdom I nterrogatory model card

2023
[43]

Haoxi Zhong, Chaojun Xiao, Cunchao Tu, Tianyang Zhang, Zhiyuan Liu, and Maosong Sun. 2020. How does nlp benefit legal system: A summary of legal artificial intelligence. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5218--5230

2020
[44]

Zhi Zhou, Jiang-Xin Shi, Peng-Xiao Song, Xiao-Wen Yang, Yi-Xuan Jin, Lan-Zhe Guo, and Yu-Feng Li. 2024. Lawgpt: A chinese legal knowledge-enhanced large language model. arXiv preprint arXiv:2406.04614

2024 arXiv
[45]

Yaqi Zhu and Feng Xiao. 2020. The relationship between appeal requests and the method of second-instance rulings: Focusing on article 170, paragraph 1 of the *civil procedure law*. Legal Application, (11):126--133

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Bodenheimer, Zhenlai Deng, and Jingwu Ji

E. Bodenheimer, Zhenlai Deng, and Jingwu Ji. 1987. Jurisprudence: Philosophy of Law and Its Methods. Huaxia Publishing House

work page 1987

[4] [4]

Marius B \"u ttner and Ivan Habernal. 2024. https://aclanthology.org/2024.eacl-long.122/ Answering legal questions from laymen in G erman civil law system . In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2015--2027, St. Julian ' s, Malta. Association for Computa...

2024

[5] [5]

Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, and Ion Androutsopoulos. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.261 LEGAL - BERT : The muppets straight out of law school . In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2898--2904, Online. Association for Computational Linguistics

doi:10.18653/v1/2020.findings-emnlp.261 2020

[6] [6]

Junyun Cui, Xiaoyu Shen, Feiping Nie, Zheng Wang, Jinglong Wang, and Yulong Chen. 2022. https://doi.org/10.48550/arXiv.2204.04859 A survey on legal judgment prediction: Datasets, metrics, models and challenges . CoRR, abs/2204.04859

arXiv doi:10.48550/arxiv.2204.04859 2022

[7] [7]

Junyun Cui, Xiaoyu Shen, and Shaochun Wen. 2023. A survey on legal judgment prediction: Datasets, metrics, models and challenges. IEEE Access, 11:102050--102071

2023

[8] [8]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long a...

doi:10.18653/v1/n19-1423 2019