Pith. sign in

REVIEW 3 major objections 8 minor 2 cited by

TransLaw: A Large-Scale Dataset and Multi-Agent Benchmark Simulating Professional Translation of Hong Kong Case Law

T0 review · 3 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TransLaw, a seven-agent LLM pipeline with glossary lookup and iterative review, beats single-model translation for Hong Kong case law on all 13 models tested and ships the first sentence-aligned bilingual corpus for the domain.

desk verdict A genuinely new parallel corpus and domain benchmark for HK case law; the multi-agent gains are consistent, but the headline convergence claim is under-supported and needs one more analysis. read the letter →

arxiv 2507.00875 v3 pith:6OLBHMDZ submitted 2025-07-01 cs.CL cs.HCcs.MA

Xi Xuan , Chunyu Kit This is my paper
classification cs.CLcs.HCcs.MA
keywords HongKongcaselawlegalmachinetranslationmulti-agentLLMsystemretrieval-augmentedgenerationparallelcorpusTraditionalChineseCourtofFinalAppealevaluation
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 claims that translating Hong Kong Court of Final Appeal judgments from English into Traditional Chinese—a constitutional requirement under the Basic Law that human effort alone cannot scale—can be made substantially more accurate by splitting the task among specialized large-language-model agents rather than asking a single model to do everything. It introduces HKCFA Judgment 97-22, the first sentence-aligned parallel corpus for Hong Kong case law, built from 344 professionally translated judgments spanning 1997 to 2022, and reports that its TransLaw framework outperforms single-agent translation on all 13 models tested, with GPT-4o rising from 69.42 to 85.12 on xCOMET-XL and every system converging within three review iterations. A human evaluation by 10 certified legal translators using the proposed Legal ACS metric confirms gains in legal-semantic accuracy, while showing that human experts still exceed the system in coherence and stylistic naturalness. If the claims hold, the framework offers an assistive pipeline about three orders of magnitude cheaper than professional human translation of the same judgment (about US$0.35 versus US$1,390 in API cost).

What carries the argument

The load-bearing mechanism is the seven-agent collective $C=\{A_{\mathrm{Com}}, A_{\mathrm{Term}}, A_{\mathrm{Trans}}, A_{\mathrm{Align}}, A_{\mathrm{TermR}}, A_{\mathrm{Cita}}, A_{\mathrm{StyleP}}\}$, organised into a command module, an execution module, and a review module. The iteration law is $\hat{s}_i^{(k+1)} = \hat{s}_i^{(k)} \oplus \Psi(F_i^{(k)})$, where $F_i^{(k)}$ is the feedback from review round $k$ and $\Psi$ turns that feedback into textual edits; the loop stops when $F_i^{(k)} = \varnothing$ or the threshold $K = 3$ is reached. Two components carry most of the weight: retrieval-augmented generation over the government glossary, which fixes legally binding renderings the translator agent may not paraphrase, and the four-way expert review that re-checks the draft against the source sentence, the glossary, the citation conventions, and the register of a court judgment. On the human-evaluation side the paper proposes the Legal ACS metric $I = 0.6A + 0.3C + 0.1S$, weighting accuracy of legal meaning above coherence and style.

What would settle it

Take a random sample of about 200 of the 11,099 sentence pairs in HKCFA Judgment 97-22 and have two bilingual legal professionals independently judge whether each English sentence and its Traditional Chinese counterpart are complete, faithful renderings of one another; if more than a few percent are misaligned, truncated, or paired across non-corresponding paragraphs, the gold references and every score derived from them are in doubt.

Watch

Extended reading notes

Core claim

TransLaw's central claim is that professional-grade legal translation decomposes cleanly into roles that separate LLM agents can play, and that this role-split with iterative review beats any single model translating alone. The system chains a command agent that segments sentences and keeps a memory of prior finalized translations; a terminology agent that retrieves candidates from the official Combined DOJ Glossaries of Legal Terms; a translator agent that must use the retrieved renderings verbatim while preserving Chinese syntax; and four review agents that check semantic alignment, glossary compliance, citation format, and judicial tone. Drafts are revised against reviewer feedback through the update rule $\hat{s}_i^{(k+1)} = \hat{s}_i^{(k)} \oplus \Psi(F_i^{(k)})$ until feedback is empty or the third iteration, which the paper reports as the convergence point for all 13 models. The ablation attributes the largest score drops to removing the translator and command agents, then to removing the glossary and its retrieval step, and the human evaluation shows the framework gains on accuracy while still trailing professional translators on coherence and style.

Load-bearing premise

The whole benchmark stands on the assumption that the 11,099 sentence pairs are truly parallel because they were cut from official bilingual HTML paragraphs; the paper reports no alignment error rate and no manual check of the pairs, so if the official paragraph boundaries do not correspond one-to-one to translation units, the gold references and every score built on them are contaminated.

Editorial extensions

If this is right

  • If the reported gains hold, any of the 13 tested models—including 6B-parameter open models—becomes a materially better legal translator when embedded in the role-split loop, so the benefit is not tied to one commercial API.
  • Because all systems converge within three iterations, the multi-agent pipeline is a bounded, practical process rather than an open-ended debate loop.
  • The released corpus gives the first sentence-aligned English–Traditional Chinese benchmark for Hong Kong case law, which future legal machine-translation work can use as gold data.
  • The ablation shows the official glossary database and its retrieval step are not replaceable by the models' parametric knowledge, implying that jurisdiction-specific terminology resources are the key reusable asset.
  • At roughly US$0.35 of API cost per judgment versus about US$1,390 for professional translation of the same text, the framework changes the economics of the constitutional bilingual mandate.

Reading between the lines

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

  • A direct test of what drives the gain would be to run the same role prompts sequentially inside one model with no separate agent instances and no glossary retrieval; if the gap narrows, role decomposition itself matters less than the glossary lookup and the review prompts.
  • The corpus recipe—bilingual judgments already published as paired official HTML—should extend to other courts in the Hong Kong hierarchy and to other bilingual common-law jurisdictions, provided their official markup has the same one-to-one paragraph correspondence.
  • The two nearly deterministic review dimensions (citation format and glossary compliance) could be enforced by rule-based code, leaving the alignment and style agents as the only LLM reviewers and cutting API cost further.
  • The human evaluation's finding that translators still lead on coherence and style suggests the near-term endpoint is human-in-the-loop editing of TransLaw output rather than fully automatic publication.
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.

Referee Report

3 major / 8 minor

Summary. The paper introduces HKCFA Judgment 97-22, a sentence-aligned English-to-Traditional-Chinese parallel corpus of 344 Hong Kong Court of Final Appeal judgments (11,099 sentence pairs; about 2.1M tokens), and proposes TransLaw, a multi-agent LLM framework that decomposes legal translation into terminology retrieval (with RAG from the DOJ glossary), sentence translation, and a four-agent review loop covering semantic alignment, terminology, citation, and style, with iterative feedback. The evaluation benchmarks 13 LLMs with xCOMET-XL and wmt22-unite-da, reports ablations, and includes human evaluation by 10 certified translators using a proposed Legal ACS metric, together with cost estimates. The paper claims that TransLaw significantly outperforms single-agent baselines across all evaluated models and converges within three iterations.

Significance. The corpus fills a genuine resource gap for a high-stakes legal translation domain, and the multi-agent framework is a plausible assistive pipeline rather than a fully autonomous replacement. The automated evaluation is a strength: 13 models, bootstrap confidence intervals, independent learned metrics, and systematic ablations support the central quality claim. The human evaluation uses domain experts, which is appropriate, and the public release of data and code is valuable. However, the convergence claim, the validation of the corpus alignment, and the reporting of the human evaluation all need strengthening before the paper is fully convincing.

major comments (3)
  1. [Section 3.3, Eq. (2) and Section 4.1 (Iteration Efficiency)] The stopping rule is 'feedback is empty OR the iteration threshold K is reached', but K is never reported. Table 4's Iter column therefore cannot establish convergence: if K=3, a sentence that still has non-empty feedback at round 3 is frozen and counted as converged. Please report the value of K, the number or proportion of sentences that terminate via the empty-feedback condition versus via the cap, and revise the abstract's 'convergence within 3 iterations' claim accordingly.
  2. [Section 3.1] No alignment validation is reported. The claim that relying on official government HTML paragraph structure yields 'high-quality alignment' needs direct support, since every automatic and human score is computed against this corpus as gold reference. Please provide a validation protocol (e.g., human review of a random sample with an alignment error rate) and describe how footnotes, citations, headers, and list items are handled.
  3. [Section 4.2 and Appendix H] The human evaluation is based on a single judgment (FACC 1/2021) and reports no inter-annotator agreement, no per-annotator score distribution, and no statistical test of the differences shown in Figure 2. In addition, the scoring protocol is unclear: Section 4.2 describes 0-10 scores on three dimensions (A, C, S), while Appendix H describes a single point-deduction system. Please report inter-annotator agreement (e.g., Krippendorff's alpha), clarify how the Appendix H guidelines map to the ACS dimensions, and state how many annotators scored each segment.
minor comments (8)
  1. [Appendix E, Table 7] Table 7 lists ChatGLM-6B and Baichuan-7B-Base, which do not appear in Table 2; please align the model inventory with the models actually evaluated.
  2. [Section 4.1, first bullet] The claim that all open-source LLMs perform slightly under closed-source ones is contradicted by Table 2, where DeepSeek MoE-16B and DeepSeek-R1-Distill-Qwen-32B outperform ChatGPT; please rephrase.
  3. [Section 3.3, Eq. (2)] The mapping function Ψ(·) is introduced but never formally defined; the prompt templates give examples, but a precise definition or a pointer to the implementation would improve reproducibility.
  4. [Section 3.3] The notation for review feedback is inconsistent: Eq. (2) uses F_i^(k), while Eq. (11) uses δ_i^(k); please unify the notation.
  5. [Appendix B] The appendix title contains a typo: 'HKACFA' should likely be 'HKCFA'.
  6. [References] The reference to Sin et al. (2026) is incomplete, with no venue or publisher details beyond 'University of Hong Kong, China'.
  7. [Section 7 (Ethics Statement)] The ethics statement says the dataset contains no personal or sensitive information, but the included judgment text contains personal names (e.g., 'Lai Chee Ying') and case details; please clarify whether 'personal information' means regulated personal data and how the public-source nature of the data is handled.
  8. [Section 5 (Cost Analysis)] The cost comparison mixes API costs with human proofreading costs at a rate ($0.04/word) that is not sourced to certified legal translators; please clarify the assumptions about who performs the editing and whether the rate reflects professional legal revision.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the claimed gains are evaluated with external learned metrics and independent certified translators, and the self-citation is not load-bearing.

full rationale

The paper's derivation chain is largely self-contained and externally grounded. The dataset is constructed from official bilingual HKCFA judgments using the structure of government HTML, with a stated alignment rationale rather than a fitted parameter. TransLaw's components (ACom, ATerm, ATrans, AAlign, ATermR, ACita, AStyleP) are prompted role agents, not parameters learned from the test set. Automated evaluation uses xCOMET-XL and wmt22-unite-da, which are external learned metrics not trained on HKCFA data, and the human evaluation uses 10 certified legal translators with an ACS metric whose weights (0.6/0.3/0.1) are expert-chosen and checked against alternative weightings. The only self-citation (Sin et al., 2026) supports a background claim about the importance of large-scale legal translation and is not load-bearing. One caveat is that the iteration cap K in Eq. (2) is never reported, so the abstract's 'convergence within 3 iterations' is under-specified and could in principle reflect the cap rather than observed empty feedback; however, the paper does not state K=3, so this is a verifiability concern about a secondary claim, not a demonstrated circular reduction. The primary quality gains in Tables 2 and 3 are independent of this concern.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical or formal entities are posited. The seven agent roles are software components implemented through prompts, and the Legal ACS metric is a weighted composite of human scores; both have observable implementations and are not independent postulates. The ACS metric has no external falsifiable handle, so it is not treated as supporting independent evidence.

free parameters (2)
  • Legal ACS weights (alpha, beta, gamma) = 0.6, 0.3, 0.1
    Chosen 'based on the experience and recommendation of domain experts' (Section 4.2) to weight accuracy, coherence, and style in the human-evaluation metric. Not fitted to data, but hand-picked; Figure 2 shows alternative weights give minor score changes.
  • Iteration bound K = 3 (implicit)
    The refinement loop in Eq. (2) terminates when feedback is empty 'or the iteration threshold K is reached' (Section 3.3). The paper reports convergence within 3 iterations without stating K independently, so the convergence claim may be partly by construction.
assumptions (5)
  • domain assumption The Combined DOJ Glossaries of Legal Terms is the authoritative source for HK legal terminology.
    Used by ATerm and ATermR in Equations (3) and (8) and throughout the framework; if the glossary is incomplete or outdated, terminology fidelity is compromised.
  • domain assumption The official English and Chinese versions of HKCFA judgments are professionally translated and can serve as gold references.
    The dataset and the human-evaluation 'gold translation' both assume the official court translations are correct (Sections 3.1 and 4.2).
  • domain assumption xCOMET-XL and wmt22-unite-da scores are valid predictors of legal translation quality for EN to ZH-Trad.
    All automated conclusions in Section 4.1 rely on these learned metrics without a dedicated validation on legal Traditional Chinese.
  • domain assumption Paragraph structure in official judgment HTML yields correct sentence-level alignment.
    Section 3.1 claims 'high-quality alignment' from HTML structure; no error analysis is provided.
  • domain assumption LLM agents give reliable review feedback rather than hallucinated corrections.
    The four review agents are LLMs with no ground-truth check on their feedback; the paper assumes their error reports improve translations (Section 3.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of TransLaw: A Large-Scale Dataset and Multi-Agent Benchmark Simulating Professional Translation of Hong Kong Case Law." pith.science (2026). https://pith.science/paper/6OLBHMDZ

@misc{pith2026250700875,
  author       = {Pith},
  title        = {Pith review of: TransLaw: A Large-Scale Dataset and Multi-Agent Benchmark Simulating Professional Translation of Hong Kong Case Law},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6OLBHMDZ}},
  note         = {Machine review of arXiv:2507.00875}
}
read the original abstract

Translating Hong Kong Court Judgments from English to Traditional Chinese is mandated by Articles 8-9 of the Basic Law, yet remains constrained by a shortage of parallel resources and rigorous demands on legal terminology, citation format, and judicial style. We introduce HKCFA Judgment 97-22, the first large-scale sentence-aligned parallel corpus for HK case law, comprising 344 professionally translated judgments (11,099 sentence pairs; 2.1M tokens) spanning 1997-2022. Building on this resource, we propose TransLaw, a multi-agent framework that decomposes translation into word-level expression, sentence-level translation, and multidimensional review, integrating a specialized Hong Kong legal glossary database, Retrieval-Augmented Generation, and iterative feedback, with four-dimensional expert review covering semantic alignment, terminology, citation, and style. Benchmarking 13 open-source and commercial LLMs, we demonstrate that TransLaw significantly outperforms single-agent baselines across all evaluated models, with convergence within 3 iterations. Human evaluation by 10 certified legal translators using our proposed Legal ACS metric confirms gains in legal-semantic accuracy, while showing that TransLaw still trails human experts in stylistic naturalness. The dataset and benchmark code are available at https://github.com/xuanxixi/TransLaw.

Figures

Figures reproduced from arXiv: 2507.00875 by the authors.

Figure 1
Figure 1. The overall architecture of TransLaw. The framework consists of three collaborative modules: (1) the Translation Command Module (MCom), where ACom coordinates the global workflow; (2) the Translation Execution Module (MExec), comprising ATerm and ATrans for legal terminology parsing and core translation; and (3) the Expert Review Module (MRev), which integrates AAlign, ATermR, ACita, and AStyleP for multi-dimensiona… view at source ↗
Figure 2
Figure 2. Human Evaluation Results. Performance of the three systems across three dimensions (top) and different weighting schemes (bottom). segment lengths were controlled, with the maximum length reaching 234 English words (290 tokens) and 414 Chinese words (580 tokens). Anonymized evaluation tables contain￾ing segment/sentence IDs, source texts, and system outputs were assessed by 10 certified professional legal translator… view at source ↗
Figure 3
Figure 3. Example of the first page of a bilingual HKCFA judgment (Case No. FACC 1/2021), comparing the source text and target translation. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Prompt template for Senior Translation Project Manager. It covers segmentation, context memory, iterative feedback integration, and final result aggregation. Prompt Template for Legal Terminology Agent (ATerm) Role: You are a HK Legal Terminologist working for the Depa…
Figure 5
Figure 5. Figure 5: Prompt template for HK Legal Terminologist, it ensures initial terms are retrieved from official glossary. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Prompt template for Senior Court Translator, it integrates mandatory terminology into the target syntactic structure to generate the translation candidate. Prompt Template for Semantic Alignment Agent (AAlign) Role: You are a Senior Legal Reviser responsible for compar…
Figure 7
Figure 7. Figure 7: Prompt template for Senior Legal Reviser, it performs a bilingual check to detect semantic errors or omissions. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Prompt template for Chief Terminologist, it validates lexical accuracy against the authoritative government glossary. Prompt Template for Legal Citation Agent (ACita) Role: You are a Professional Legal Editor specializing in HKCFA judgment citation standards. Task: For…
Figure 9
Figure 9. Figure 9: Prompt template for Professional Legal Editor, it enforces rigid citation formatting standards for cases and legislation. Prompt for Stylistic Fidelity Polishing Agent (AStyleP) Role: You are a Senior Judicial Editor responsible for refining the text to uphold the auth…
Figure 10
Figure 10. Figure 10: Prompt template for Senior Judicial Editor, it refines the linguistic output to ensure judicial authority and fluency. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: The detailed evaluation guidelines provided to human experts. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Multilingual Source Tracing of Speech Deepfakes: A First Benchmark

    eess.AS 2025-08 conditional novelty 6.0 of 10

    The first multilingual source-tracing benchmark for speech deepfakes, showing LFCC-ECAPA-TDNN generalizes best across languages.

  2. Fake-Mamba: Real-Time Speech Deepfake Detection Using Bidirectional Mamba as Self-Attention's Alternative

    eess.AS 2025-08 unverdicted novelty 5.0 of 10

    Fake-Mamba reports EERs of 0.97%, 1.74%, and 5.85% on three speech deepfake benchmarks, but the provided full text is an unrelated paper, so the claims cannot be verified.

Reference graph

Works this paper leans on

27 extracted references · 18 canonical work pages · cited by 2 Pith papers

  1. [1]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    arXiv

  2. [8]

    Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts lan- guage models.arXiv preprint arXiv:2401.06066,

    Dai, D., Deng, C., Zhao, C., Xu, R., Gao, H., Chen, D., Li, J., Zeng, W., Yu, X., Wu, Y ., et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts lan- guage models.arXiv preprint arXiv:2401.06066,

    arXiv

  3. [9]

    Tear: Improv- ing llm-based machine translation with systematic self- refinement

    Feng, Z., Zhang, Y ., Li, H., Wu, B., Liao, J., Liu, W., Lang, J., Feng, Y ., Wu, J., and Liu, Z. Tear: Improv- ing llm-based machine translation with systematic self- refinement. InFindings of the Association for Computa- tional Linguistics: NAACL 2025, pp. 3922–3938,

    work page 2025

  4. [10]

    M., Rei, R., Stigt, D

    9 ICML 2026 - AI for Law Guerreiro, N. M., Rei, R., Stigt, D. v., Coheur, L., Colombo, P., and Martins, A. F. xcomet: Transparent machine translation evaluation through fine-grained error detec- tion.Transactions of the Association for Computational Linguistics, 12:979–995,

    work page 2026

  5. [11]

    Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

  6. [12]

    P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., et al

    Hurst, A., Lerer, A., Goucher, A. P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    arXiv

  7. [13]

    Jones Jr, D

    https://github.com/openai/ tiktoken. Jones Jr, D. A. A leg to stand on-post-1997 hong kong courts as a constraint on prc abridgment of individual rights and local autonomy.Yale J. Int’l L., 12:250,

    work page 1997

  8. [15]

    doi: 10.18653/v1/2024.emnlp-main.992

    Association for Computational Linguistics. doi: 10.18653/v1/2024.emnlp-main.992. URL https: //aclanthology.org/2024.emnlp-main.992/. Liu, L., Wang, L., Niu, W., and Hua, S. Dynamic sanc- tioning mechanism for cooperative multi-agent systems. Expert Systems with Applications, 296:128873,

    doi:10.18653/v1/2024.emnlp-main.992 2024

Show all 27 references
  1. [19]

    Examining inter-consistency of large language models collaboration: An in-depth analysis via debate

    Xiong, K., Ding, X., Cao, Y ., Liu, T., and Qin, B. Examining inter-consistency of large language models collaboration: An in-depth analysis via debate. InFindings of the Asso- ciation for Computational Linguistics: EMNLP 2023, pp. 7572–7590,

    2023

  2. [20]

    Baichuan 2: Open large-scale language models.arXiv preprint arXiv:2309.10305,

    Yang, A., Xiao, B., Wang, B., Zhang, B., Bian, C., Yin, C., Lv, C., Pan, D., Wang, D., Yan, D., et al. Baichuan 2: Open large-scale language models.arXiv preprint arXiv:2309.10305,

    arXiv

  3. [21]

    Use of AI Assistants We used Claude Opus 4.7 and Sonnet 4.6 for coding, shortening texts and editing LaTeX more efficiently

    11 ICML 2026 - AI for Law A. Use of AI Assistants We used Claude Opus 4.7 and Sonnet 4.6 for coding, shortening texts and editing LaTeX more efficiently. B. HKACFA Judgment Example (Case No. FACC 1/2021) FACC No. 1 of 2021

    2026

  4. [23]

    BETWEEN HKSAR Appellant and LAI CHEE YING (黎智英) Respondent Before: Chief Justice Cheung, Mr Justice Ribeiro PJ, Mr Justice Fok PJ, Mr Justice Chan NPJ and Mr Justice Stock NPJ Dates of Hearing: 1 February 2021 Date of Judgment: 9 February 2021 J U D G M E N T The Court:

    2021

  5. [24]

    (a)English Judgment (Source) A B C D E F G H I J K L M N O P Q R S T U V A B C D E F G H I J K L M N O P Q R S T U V [Chinese Translation - 中譯本] FACC 1/2021

    2021

  6. [25]

    FACC 1/2021), comparing the source text and target translation

    HKCFA 3 香港特別行政區 終審法院 終院刑事上訴案件2021年第1號 （原高院刑事雜項案件2020年第727及738號） 上訴人 香港特別行政區 對 答辯人 黎智英 主審法官： 終審法院首席法官張舉能 終審法院常任法官李義 終審法院常任法官霍兆剛 終審法院非常任法官陳兆愷 終審法院非常任法官司徒敬 聆訊日期： 2021年2月1日 判決日期： 2021年2月9日 (b)Chinese Translation (Target) Figure 3.Example of the first page of a bilingual HKCFA judg...

    2021

  7. [26]

    task_mode

    or legislative references (e.g., Section 9 of the Theft Ordinance (Cap. 210)). -Note 2:Preserve the logical flow. Do not break bullet points if they constitute a single semantic unit. Task 2: Context ManagementMaintain a record of previously finalized translations to ensure te...

    2026

  8. [27]

    Cap." and

    HKCFA 8). -Legislation:Ensure correct use of "Cap." and "Section". (e.g., Theft Ordinance (Cap. 210)). Input Data (JSON):{ "candidate": "{{draft}}" } Output Format (JSON):{ "status": "PASS", "feedback": null } Figure 9.Prompt template for Professional Legal Editor, it enforces...

  9. [1999]

    V ., Guerreiro, N

    Treviso, M. V ., Guerreiro, N. M., Agrawal, S., Rei, R., Pom- bal, J., Vaz, T., Wu, H., Silva, B., Van Stigt, D., and Martins, A. F. xtower: A multilingual llm for explaining and correcting translation errors. InFindings of the Asso- ciation for Computational Linguistics: EMNL...

    2024

  10. [2002]

    and He, L

    Cheng, L. and He, L. Revisiting judgment translation in hong kong.Semiotica, 2016(209):59–75,

    2016

  11. [2012]

    Llm-based translation inference with iterative bilingual understanding

    Chen, A., Chen, K., Xiang, Y ., Bai, X., Yang, M., Feng, Y ., Zhao, T., and Zhang, M. Llm-based translation inference with iterative bilingual understanding. InFindings of the Association for Computational Linguistics: ACL 2025, pp. 16886–16902,

    2025

  12. [2016]

    Chatlaw: A multi-agent col- laborative legal assistant with knowledge graph enhanced mixture-of-experts large language model.arXiv preprint arXiv:2306.16092,

    Cui, J., Ning, M., Li, Z., Chen, B., Yan, Y ., Li, H., Ling, B., Tian, Y ., and Yuan, L. Chatlaw: A multi-agent col- laborative legal assistant with knowledge graph enhanced mixture-of-experts large language model.arXiv preprint arXiv:2306.16092,

    arXiv

  13. [2020]

    A., Humbel, C., Gösken, C., Tanzi, L., Lüthi, T., et al

    Niklaus, J., Merane, J., Nenadic, L., Ahmadi, S., Gao, Y ., Chevalley, C. A., Humbel, C., Gösken, C., Tanzi, L., Lüthi, T., et al. Swiltra-bench: The swiss legal translation benchmark.arXiv preprint arXiv:2503.01372,

    arXiv

  14. [2021]

    1 OF 2021 (CRIMINAL) (ON APPEAL FROM HCCP NOS

    HKCFA 3 IN THE COURT OF FINAL APPEAL OF THE HONG KONG SPECIAL ADMINISTRATIVE REGION FINAL APPEAL NO. 1 OF 2021 (CRIMINAL) (ON APPEAL FROM HCCP NOS. 727 AND 738 OF

    2021

  15. [2022]

    International and supranational law in translation: From multilingual lawmaking to adjudication

    10 ICML 2026 - AI for Law Prieto Ramos, F. International and supranational law in translation: From multilingual lawmaking to adjudication. The Translator, 20(3):313–331,

    2026

  16. [2023]

    Qwen technical report.arXiv preprint arXiv:2309.16609,

    Bai, J., Bai, S., Chu, Y ., Cui, Z., Dang, K., Deng, X., Fan, Y ., Ge, W., Han, Y ., Huang, F., et al. Qwen technical report.arXiv preprint arXiv:2309.16609,

    arXiv

  17. [2024]

    Benita, J., Jaswanth, S., Bhuvaneshwar, N., Yuvaraj, R., and Narayana, Y . L. Phoenix: A conversational agent for emotional well-being and psychological support. In 2025 International Conference on Multi-Agent Systems for Collaborative Intelligence (ICMSCI), pp. 1137–1142. IEEE,

    2025

  18. [2025]

    Chen, A. H. Y . Hong kong’s legal system in the new consti- tutional order: The experience of 1997–2000. InImple- mentation of Law in the People’s Republic of China, pp. 213–245. Brill Nijhoff,

    1997

  19. [2026]

    Encouraging diver- gent thinking in large language models through multi- agent debate

    Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y ., Wang, R., Yang, Y ., Shi, S., and Tu, Z. Encouraging diver- gent thinking in large language models through multi- agent debate. In Al-Onaizan, Y ., Bansal, M., and Chen, Y .-N. (eds.),Proceedings of the 2024 Conference on Empir...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.