Pith. sign in

REVIEW 2 major objections 5 minor 35 references

BriefMe: A Legal NLP Benchmark for Assisting with Legal Briefs

T0 review · 2 major / 5 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 Today's large language models can already write persuasive brief headings, sometimes better than human lawyers, but they struggle to place missing arguments and retrieve relevant case law.

desk verdict BriefMe is a genuinely useful legal NLP benchmark, but the headline 'LLMs beat human headings' rides on an LLM judge that was also used to filter the data, so that claim needs a stronger human anchor. read the letter →

arxiv 2506.06619 v3 pith:KG3UMZI6 submitted 2025-06-07 cs.CL

Jesse Woo , Fateme Hashemi Chaleshtori , Ana Marasović , Kenneth Marino This is my paper
classification cs.CL
keywords legalNLPbriefsbenchmarkdatasetargumentsummarizationcompletioncaseretrievalLLM-as-a-judgeSupremeCourt
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 legal brief writing can be broken into three measurable assistive tasks—summarizing an argument as a heading, completing a missing argument in a brief's table of contents, and retrieving the precedent a passage should cite—and introduces a benchmark dataset built from Supreme Court briefs to test them. On the first two tasks, current large language models are already strong: GPT-4o's generated headings are rated above human-authored headings by an LLM judge, and the result holds on briefs published after the models' training cutoffs. On the harder versions of the tasks, the models lag badly: Llama-3.1-70B detects that a heading is missing but places it correctly only 18% of the time, and the best retriever finds the cited precedent in its top five only 31.5% of the time. A sympathetic reader should care because these are the skills that consume lawyer time, and the benchmark identifies which ones are ready for assistive tools and which are not.

What carries the argument

The load-bearing object is the dataset itself, built by scraping and parsing 3,753 Supreme Court merit briefs from 2017 to 2024 and aligning each section's heading to its body text. That structure supplies all three tasks at once: each heading is a gold extreme summary of its section; each table of contents is a tree of nested arguments from which one node can be masked for completion; each citation, masked with a unique identifier, defines a retrieval query against a corpus of 24,525 cited opinions. Evaluation is carried by an LLM-as-judge: o3-mini, prompted with expert guidelines on persuasive legal headings, assigns ratings from 1 to 5 that the paper argues are more reliable than the human annotators it recruited.

What would settle it

Take a random sample of, say, 100 summarization outputs from few-shot GPT-4o and the human-authored headings from the same briefs, have a panel of practicing appellate lawyers rate them blind with the same rubric, and check whether the human headings actually average below the model's; if expert agreement does not reproduce the judge's ordering, the 'beats human headings' claim fails.

Watch

Extended reading notes

Core claim

The central discovery is BRIEF ME, a benchmark that repurposes the internal structure of Supreme Court merit briefs—section headings, nested tables of contents, and masked citations—into three training and evaluation tasks. The paper's empirical claim is that LLMs are already useful for the drafting side of legal work: few-shot GPT-4o headings average 4.3/5 from the o3-mini judge, above the 4.0 of quality-filtered human headings and well above the 3.4/3.5 of unfiltered human headings, and the advantage persists on a contamination-controlled set of briefs published after February 2025. The same models are not yet reliable for the structural and evidentiary side: locating a missing heading in a realistic table-of-contents completion succeeds only 18% of the time, and no retrieval method brings the correct precedent into the top five more than 31.5% of the time, even after fine-tuning and reranking. The intended upshot is a division of labor: LLMs can draft headings, while placement and precedent retrieval remain open problems.

Load-bearing premise

The claim that GPT-4o beats human headings stands or falls with the assumption that the o3-mini judge's ratings really capture expert legal heading quality, a check done on just ten samples per task with meta-ratings by one author.

Editorial extensions

If this is right

  • Few-shot GPT-4o produces section headings rated 4.3/5 on average, above human headings, and its advantage survives on briefs published after the model's training cutoff, so heading generation is a realistic near-term assistive tool.
  • Fine-tuning and in-context examples often hurt open models on the judge's scale, so prompting a strong proprietary model currently beats training smaller ones.
  • A model can detect that a table of contents is incomplete and name the missing heading's level, but placing it correctly succeeds only 18% of the time; this is the bottleneck for realistic argument completion.
  • Case retrieval is the weakest task, with the best top-five recall around 31.5%, and models often retrieve a topically similar but not the cited case, so retrieval needs new methods before it can support brief drafting.
  • Because the same source documents yield all three tasks, BRIEF ME lets researchers study drafting and research skills on one corpus rather than stitching together datasets.

Reading between the lines

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

  • If the judge's ratings are trusted, the practical design lesson is to spend engineering effort on placement and retrieval rather than phrasing: a drafting assistant that suggests headings is near-ready, while one that decides where they go and what to cite is not.
  • The 'right topic, wrong case' failure pattern suggests the exact-match evaluation may understate real utility, because lawyers often need any supportive precedent; a natural extension is to score retrieval by whether the top results could be cited for the proposition.
  • The judge's reliability rests on ten samples per task and one author's meta-ratings, so the 'beats human headings' claim should be rechecked with a larger expert panel before it is used to justify deployment.
  • The three-task structure invites a testable composite: use the retrieval task to supply evidence for candidate completed headings, and use summarization quality as a signal that a placed heading fits; that could close the 18% placement gap.
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

2 major / 5 minor

Summary. The paper introduces BRIEFME, a legal NLP benchmark built from Supreme Court of the United States merit briefs, with three tasks: argument summarization (generating section headings), argument completion (filling a missing heading in a table of contents, in both guided and realistic settings), and case retrieval (retrieving the case cited in a brief passage). The authors describe a detailed extraction and cleaning pipeline, present dataset statistics, and benchmark a range of LLMs and retrieval models. Using an o3-mini LLM-as-a-judge, they report that current LLMs, especially GPT-4o, achieve judge ratings above human-authored headings on summarization and guided completion, while performing poorly on realistic argument completion and on retrieving the exact cited case. They also include a contamination check using briefs published after model training cutoffs.

Significance. If the evaluation is valid, BRIEFME fills a clear gap in Legal NLP by focusing on persuasive legal argumentation in briefs rather than judicial opinions. The dataset construction is careful and detailed, the contamination check with temporally separated briefs is a notable strength, and the authors have released code and data. The retrieval corpus of over 24K cases and the three-task structure should be useful to the community. However, the headline claim that LLMs 'beat human-generated headings' is not yet established because it rests on an LLM judge that was also used to filter the dataset, was validated on only ten samples per task, and was meta-evaluated by one of the authors. The retrieval evaluation also assumes a single gold case. These issues do not undermine the dataset itself but do require substantial revision before the central comparative claims can be accepted.

major comments (2)
  1. [§4.1, §4.2, Table 3, Appendix E/F] The central claim that GPT-4o few-shot 'beats human-generated headings' (4.3 vs. 4.0 on the filtered set and vs. 3.4 on the unfiltered set, Table 3) is carried entirely by the o3-mini LLM-as-judge. The same judge is used in §3.4 to filter low-quality human headings and then used to compare model outputs against human headings, so the judge both constructs the gold set and measures the outcome. More importantly, the generation prompt (Appendix F.1) and the judge rubric (Appendix F.2) are both derived from the same legal-writing guidelines; model outputs are therefore generated to satisfy the exact criteria the judge applies, while human headings were not. The external validation of the judge consists of three annotators on ten samples per task, with meta-ratings provided by one of the paper's authors (Appendix E, Figs. 15–16, Table 20). Furthermore, the authors' own error analysis (Appendix H.1/H.2) shows the judge assigning different scores (4 vs. 5) to identical or near-identical headings, so the 4.3-vs-4.0 margin falls within the judge's apparent inconsistency. To support the headline claim, the authors should either provide a substantially larger independent expert evaluation that calibrates the 4-vs-5 boundary and directly compares model vs. human headings, or soften the claim to state that models reach a level comparable to human headings according to an LLM judge.
  2. [§4.4, Table 5, Appendix H.3] The case retrieval evaluation treats the single case cited in the brief as the only relevant document. As the authors acknowledge in Appendix H.3, in 2 of 40 examined failures both models retrieved cases that could plausibly have been cited for the proposition, and in many other cases the models retrieved topically correct cases. Under a single-gold assumption, Recall@k and MRR@10 (Table 5) may substantially underestimate retrieval quality. The claim that models 'perform poorly' on case retrieval should be re-framed as 'models rarely retrieve the exact cited case,' or the authors should provide a supplementary multi-relevance evaluation, even on a small sample, to separate topical relevance from exact-match retrieval. Without this, the negative result for case retrieval is not crisply interpretable.
minor comments (5)
  1. [§4.2] The sentence 'the LLM judge scores show that sincorporating training examples' contains a typo; it should read 'since incorporating.'
  2. [Appendix I] The model name 'Qwe-n2.5-14B-Instruct' is misspelled; it should be 'Qwen-2.5-14B-Instruct.'
  3. [Appendix B, Table 7] Table 7 reports 'Mean Compression 59.90' while the text states that the dataset has a 'higher median Compression rate (∼1.3x) than Newsroom'; please clarify whether this is a ratio, a percentage, and whether the mean or median is intended, since the two figures are inconsistent in magnitude.
  4. [Table 3 caption] The notation 'o3-miniS' is used in Table 3 and later tables, but it is not defined in the caption or in the main text; please define it explicitly when it is first introduced.
  5. [§4.1] The footnote 'We refined the judge instructions using Claude' is vague; a brief description of the refinement process and whether the final instructions were re-validated on the human-judge comparison would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the same LLM judge used for filtering and evaluation is a validity limitation, not a reduction; benchmark results are not forced by construction.

full rationale

The paper's central outputs are empirical benchmark scores, not a derivation from fitted inputs. The o3-mini judge is used both to filter low-quality human headings (Section 3.4) and to evaluate model outputs (Sections 4.1-4.2), which raises a legitimate evaluation-validity concern: if the judge is biased, the human-versus-model comparison is weakened. However, this is not circularity in the sense of this analysis, because the judge is a fixed, externally prompted evaluation function rather than a parameter fitted to the data being predicted; model outputs are generated without optimizing the judge, and the unfiltered comparison (4.3 vs. 3.4) does not depend on the filtering step. The human validation in Appendix E is small and partly self-assessed, but it is external grounding rather than a reduction. There is no self-citation chain, no imported uniqueness theorem, no ansatz smuggled via citation, and no renaming of a known result. The case-retrieval results are fully independent of the judge. The paper itself flags the evaluation difficulty in Section 6, and the claims are stated with appropriate hedging. Therefore no significant circularity is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical or theoretical entities are introduced. The benchmark dataset itself is a constructed resource, not a postulated entity needing independent evidence. The main ledger entries are hand-chosen filtering thresholds and domain assumptions about legal document structure.

free parameters (6)
  • Minimum section word count = 25
    Sections shorter than 25 words are excluded from summarization (Section 3.3), a hand-chosen filter.
  • Minimum heading word count = 3
    Headings under 3 words are excluded as vague or uninformative (Section 3.3).
  • Summarization quality filter threshold = score <= 3 excluded
    Section 3.4 excludes about 43% of summarization samples rated 3 or lower by the LLM judge.
  • Completion quality filter threshold = average ToC score below 3.5 excluded
    Section 3.4 excludes about 22% of completion samples with ToC scores below 3.5.
  • LoRA fine-tuning hyperparameters = rank=8, alpha=16, dropout=0.05, lr=2e-5, batch=4, 1 epoch
    Appendix I: the fine-tuning configuration is chosen by the authors and not justified by a hyperparameter search.
  • Token limits for inputs and generations = 5K input, 50 output
    Appendix I: the 50-token generation cap is twice the average human heading length; the 5K input cap accommodates all task inputs.
assumptions (5)
  • domain assumption Section headings in legal briefs are gold-standard extreme summaries of the sections they introduce.
    The argument summarization task uses the existing heading as the reference summary (Section 3.3); if headings are not faithful summaries, task validity suffers.
  • domain assumption The table of contents of a brief encodes the logical structure of legal arguments, so removing and regenerating a heading is a meaningful test of argument reasoning.
    Argument completion constructs inputs by deleting a heading from a ToC (Section 3.3).
  • ad hoc to paper The o3-mini LLM judge, prompted with legal writing guidelines, produces valid assessments of legal heading quality.
    Section 4.1: this judge is used for filtering and evaluation; validation is limited to 10 samples per task with meta-ratings by one author.
  • domain assumption The specific case citation masked in each retrieval query is the unique relevant precedent for that section.
    Section 4.4 and Appendix H.3: retrieval is scored against the single cited case, though the error analysis concedes other cases could support the same proposition.
  • domain assumption The PDF extraction and fuzzy header-to-section matching pipeline yields accurate section boundaries.
    Appendix C.5: this is validated by random sampling only, not by exhaustive manual verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BriefMe: A Legal NLP Benchmark for Assisting with Legal Briefs." pith.science (2026). https://pith.science/paper/KG3UMZI6

@misc{pith2026250606619,
  author       = {Pith},
  title        = {Pith review of: BriefMe: A Legal NLP Benchmark for Assisting with Legal Briefs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KG3UMZI6}},
  note         = {Machine review of arXiv:2506.06619}
}
read the original abstract

A core part of legal work that has been under-explored in Legal NLP is the writing and editing of legal briefs. This requires not only a thorough understanding of the law of a jurisdiction, from judgments to statutes, but also the ability to make new arguments to try to expand the law in a new direction and make novel and creative arguments that are persuasive to judges. To capture and evaluate these legal skills in language models, we introduce BRIEFME, a new dataset focused on legal briefs. It contains three tasks for language models to assist legal professionals in writing briefs: argument summarization, argument completion, and case retrieval. In this work, we describe the creation of these tasks, analyze them, and show how current models perform. We see that today's large language models (LLMs) are already quite good at the summarization and guided completion tasks, even beating human-generated headings. Yet, they perform poorly on other tasks in our benchmark: realistic argument completion and retrieving relevant legal cases. We hope this dataset encourages more development in Legal NLP in ways that will specifically aid people in performing legal work.

Figures

Figures reproduced from arXiv: 2506.06619 by the authors.

Figure 1
Figure 1. We illustrate the typical structure of briefs and highlight specific parts that serve as input and desired [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution of case citations. Shows a long [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the extraction and cleaning pipeline for B [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (27 more)
Figure 4
Figure 4. Figure 4: An example of the table that organizes case documents in [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: This figure illustrates the structure and flow of the case retrieval data. A case that is cited multiple times [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: This figure shows the top 5 topic groupings and their word scores calculated with BERTopic. ( [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Step-by-step Argument Completion Task Data Example [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]
Figure 8
Figure 8. Figure 8: (a) Argument Summarization Task: Distribution of LLM-as-judge ratings for how well the human [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: Guidelines provided to annotators for the [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 10
Figure 10. Figure 10: Example of argument summarization annotation task in the initial annotation study. [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 11
Figure 11. Figure 11: Guidelines provided to annotators for the ar [PITH_FULL_IMAGE:figures/full_fig_p027_11.png]
Figure 13
Figure 13. Figure 13: Example of argument completion annotation task in the enhanced annotation study. [PITH_FULL_IMAGE:figures/full_fig_p028_13.png]
Figure 14
Figure 14. Figure 14: Distribution of per-sample variance among 3 human rates for 10 samples of argument summarization [PITH_FULL_IMAGE:figures/full_fig_p029_14.png]
Figure 15
Figure 15. Figure 15: The left heatmap shows human ratings for 10 [PITH_FULL_IMAGE:figures/full_fig_p029_15.png]
Figure 16
Figure 16. Figure 16: The left heatmap shows human ratings for 10 [PITH_FULL_IMAGE:figures/full_fig_p030_16.png]
Figure 17
Figure 17. Figure 17: This figure shows the regular expression patterns used to identify briefs on supremecourt.gov and clean [PITH_FULL_IMAGE:figures/full_fig_p036_17.png]
Figure 18
Figure 18. Figure 18: An example of the header arguments extracted by regular expressions and manually cleaned. [PITH_FULL_IMAGE:figures/full_fig_p036_18.png]
Figure 19
Figure 19. Figure 19: An example of the header arguments extracted Mistral. The model extracted the top level headers but [PITH_FULL_IMAGE:figures/full_fig_p036_19.png]
Figure 20
Figure 20. Figure 20: This figure shows regular expressions used to find terms that indicated the beginning of the table of [PITH_FULL_IMAGE:figures/full_fig_p037_20.png]
Figure 21
Figure 21. Figure 21: This figure shows the prompt with GPT 3.5-Turbo to assist in the argument extraction process. [PITH_FULL_IMAGE:figures/full_fig_p037_21.png]
Figure 22
Figure 22. Figure 22: Summarization Task Data Example 1. This heading captures the argument of the brief about rule of [PITH_FULL_IMAGE:figures/full_fig_p037_22.png]
Figure 23
Figure 23. Figure 23: Summarization Task Data Example 2. Headings can themselves be long. Although they may be difficult [PITH_FULL_IMAGE:figures/full_fig_p038_23.png]
Figure 24
Figure 24. Figure 24: Summarization Task Data Example 3. Short sections often correspond to headers for top level arguments, [PITH_FULL_IMAGE:figures/full_fig_p038_24.png]
Figure 25
Figure 25. Figure 25: Argument Completion Task Data Example 1. The missing argument in II.A complements the point made [PITH_FULL_IMAGE:figures/full_fig_p039_25.png]
Figure 26
Figure 26. Figure 26: Argument Completion Task Data Example 2. Depending on the structure of the argument, sibling or [PITH_FULL_IMAGE:figures/full_fig_p039_26.png]
Figure 27
Figure 27. Figure 27: Argument Completion Task Data Example 3 [PITH_FULL_IMAGE:figures/full_fig_p040_27.png]
Figure 28
Figure 28. Figure 28: Argument Completion Task Data Example 4. Shorter arguments with less context are likely more difficult [PITH_FULL_IMAGE:figures/full_fig_p040_28.png]
Figure 29
Figure 29. Figure 29: Citation Retrieval Task Data Example 1 [PITH_FULL_IMAGE:figures/full_fig_p041_29.png]
Figure 30
Figure 30. Figure 30: Citation Retrieval Task Data Example 2. The eyecite annotate_citations() function works reasonably [PITH_FULL_IMAGE:figures/full_fig_p041_30.png]
Figure 31
Figure 31. Figure 31: Citation Retrieval Task Data Example 3: The retrieval dataset only annotates citations to court cases. [PITH_FULL_IMAGE:figures/full_fig_p042_31.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 29 canonical work pages

  1. [1]

    - Directly correspond to the questions presented to the court

    **Major Headings** - Present independent grounds for relief. - Directly correspond to the questions presented to the court. - Must progress logically without redundancy

    work page

  2. [2]

    - Include specific legal reasoning, claim elements, or analytical factors

    **Minor Headings** - Develop and support the arguments of their respec- tive major headings. - Include specific legal reasoning, claim elements, or analytical factors. - Should comprehensively build the case for the parent major heading

    work page

  3. [3]

    - Offer additional support and context for the argu- ments made at the minor level

    **Subheadings** - Provide further detail under minor headings. - Offer additional support and context for the argu- ments made at the minor level. ## Assessment Framework You might decide to add one or more headings at these levels: major, minor, or subheading. Here’s how you can approach determining if adding headings is needed or not, and in case it is ...

    work page 2022

  4. [4]

    The Reasoning section and the Fact section are the basis of the court’s decision

    Decision: The Decision section is the specific re- sponse given by the court to the legal dispute based on the key facts of the case. The Reasoning section and the Fact section are the basis of the court’s decision

    work page

  5. [5]

    Start of Procedure

    Tail: The Tails section introduces the basic infor- mation about the court, the judge, etc. The legal document is broken down into smaller units (sentences). Your task is to carefully identify the units that mark the beginning of each segment. Keep in mind that the segments must follow this specific sequence: Procedure, Fact, Reasoning, Decision, and Tail...

    work page

  6. [7]

    In Proceedings of the Natural Le- gal Language Processing Workshop 2021, pages 63– 75

    Multi-granular legal topic classification on greek legislation. In Proceedings of the Natural Le- gal Language Processing Workshop 2021, pages 63– 75. Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. BLEU: a method for automatic eval- uation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computa...

    work page 2021

  7. [8]

    CoRR, abs/2105.05686

    Yes, BM25 is a strong baseline for legal case retrieval. CoRR, abs/2105.05686. Fabio Salvi, Manuel H. Ribeiro, Riccardo Gallotti, and Robert West. 2025. On the conversational persua- siveness of large language models: A randomized controlled trial. Nature Human Behaviour. Advance online publication. Dan Schweitzer. 2017. U.S. supreme court brief writing s...

    arXiv 2025

  8. [13]

    - If no → The ToC is complete as is

    Would adding headings (at any level) improve the ToC’s effectiveness? - If yes → Some headings are missing, and you need to report the level of the heading as well. - If no → The ToC is complete as is

    work page

Show all 35 references
  1. [14]

    (In case the answer to the previous question is yes) Where could I add headings (at any level) that improve the ToC’s effectiveness? To confirm your choice for each heading you decide to add, check the following: - Does it highlight a missing but important argument? - Does it ...

  2. [15]

    binary_verdict

    Now that the locations are identified, the next step is generating a recommendation for the missing head- ings. Consider the following points when generating each of the new headings: - What purpose does this new heading serve? - How does this heading contribute to the overall...

  3. [16]

    E.g., ’Plaintiff’, ’Defendant’, ’On appeal’, ’procedural pos- ture’

    Procedure: The Procedure section introduces the parties’ information and procedural posture. E.g., ’Plaintiff’, ’Defendant’, ’On appeal’, ’procedural pos- ture’

  4. [17]

    Fact: The Fact section is a description of the parties’ arguments, evidence, and basic events

  5. [18]

    In Reasoning, the judge explains the reasons for the application of the rules

    Reasoning: The Reasoning component is the pro- cess where the court selects the rules and applies them to the facts. In Reasoning, the judge explains the reasons for the application of the rules. In other words, the events that are relevant to the application of the rules, i.e...

  6. [21]

    inal activity after his first offense is therefore an entirely irrelevant inquiry

    They observed the following trends. There were 17 of 30 analyzed generations rated a 1 by the judge. 11/17 were incomplete sentences (e.g. “inal activity after his first offense is therefore an entirely irrelevant inquiry.”). 2/17 were snippets of punctuation (e.g. “aa.”). 4/1...

    2009

  7. [22]

    Subjects

    “Subjects” of International Law Typically Possess Powers As Well as Obligations

  8. [23]

    Subjects

    Recognizing Corporations As “Subjects” Of International Law Is Perceived To Compromise State Sovereignty

  9. [24]

    The Lack Of International Law Consensus Regarding Corporate Liability Forecloses Such Liability Under The ATS C

    [MISSING] B. The Lack Of International Law Consensus Regarding Corporate Liability Forecloses Such Liability Under The ATS C. To The Extent That Corporate Liability Poses A Domestic Law Question, It Is One That Congress Must Answer III. DOMESTIC CORPORATE OVERSIGHT OF OVERSEAS...

  10. [25]

    Text of section 1182(a)(2)

  11. [26]

    Surrounding provisions in section 1182(a)

  12. [27]

    Renders The Alien Inadmissible

    Structure of the INA. B. The two-part structure of the stop-time rule demonstrates that Petitioner is correct. C. The purpose and history of the stop-time rule reinforce that Petitioner’s position is correct. II. Alternatively, If An Alien Is Capable Of Being Charged With Inad...

  13. [28]

    The Proposed Twitter/Facebook Standard

  14. [29]

    THE COURT SHOULD NOT ADOPT ANY OF THE NEW LEGAL RULES AND REQUIREMENTS PROPOSED BY DEFENDANTS OR THE UNITED STATES

    The Proposed Government Standard II. THE COURT SHOULD NOT ADOPT ANY OF THE NEW LEGAL RULES AND REQUIREMENTS PROPOSED BY DEFENDANTS OR THE UNITED STATES

  15. [30]

    Knowledge of Accounts or Postings Connected To The Reina Attack

  16. [31]

    Knowledge of Accounts or Postings Used for Particular Attacks

  17. [32]

    Knowledge of Substantial Assistance

  18. [33]

    Special Standard for Remoteness

  19. [34]

    Special Standard for Widely Available, Ordinary Services

  20. [35]

    THE COMPLAINT PLAUSIBLY ALLEGED THE DEFENDANTS KNOWINGLY ASSISTED ISIS’S TERRORIST ACTIVITIES A

    Requirement of Direct Knowledge III. THE COMPLAINT PLAUSIBLY ALLEGED THE DEFENDANTS KNOWINGLY ASSISTED ISIS’S TERRORIST ACTIVITIES A. The Allegations of The Complaint B. Defendants’ Asserted Lack of More Specific Knowledge Missing Heading: Special Standard for Routine Services...

  21. [1264]

    Antoine Louis and Gerasimos Spanakis

    Association for Computing Machinery. Antoine Louis and Gerasimos Spanakis. 2022. A statu- tory article retrieval dataset in french. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6789–6803, Dublin, Irel...

    2022 arXiv

  22. [2012]

    In Proceedings of the Workshop on Innovative Hybrid Approaches to the Processing of Textual Data, pages 115–123, Avignon, France

    Combining different summarization tech- niques for legal text. In Proceedings of the Workshop on Innovative Hybrid Approaches to the Processing of Textual Data, pages 115–123, Avignon, France. Association for Computational Linguistics. Mingqi Gao, Xinyu Hu, Jie Ruan, Xiao Pu, ...

    2024 arXiv

  23. [2018]

    In Proceedings of the 2018 Conference on Empirical Methods in Natural Lan- guage Processing, pages 1797–1807, Brussels, Bel- gium

    Don‘t give me the details, just the summary! topic-aware convolutional neural networks for ex- treme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Lan- guage Processing, pages 1797–1807, Brussels, Bel- gium. Association for Computational ...

    2018

  24. [2019]

    How well does the summary capture the key points of the article?

    CAIL2019-SCM: A dataset of similar case matching in legal domain. CoRR, abs/1911.08962. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayi- heng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Ji...

    2018 arXiv

  25. [2020]

    volume abs/2010.02559

    LEGAL-BERT: the muppets straight out of law school. volume abs/2010.02559. Ilias Chalkidis, Manos Fergadiotis, Prodromos Malaka- siotis, and Ion Androutsopoulos. 2019. Large-scale multi-label text classification on EU legislation. In Proceedings of the 57th Conference of the A...

    2010 arXiv

  26. [2021]

    In Proceedings of the Natural Legal Language Processing Workshop 2021, pages 19–35

    Swiss-judgment-prediction: A multilingual le- gal judgment prediction benchmark. In Proceedings of the Natural Legal Language Processing Workshop 2021, pages 19–35. Rodrigo Nogueira and Kyunghyun Cho. 2019. Passage re-ranking with BERT. CoRR, abs/1901.04085. Christos Papalouka...

    2021 arXiv

  27. [2023]

    In Proceedings of the 46th International ACM SIGIR Conference on Re- search and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023, pages 1035–1044

    SAILER: structure-aware pre-trained language model for legal case retrieval. In Proceedings of the 46th International ACM SIGIR Conference on Re- search and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023, pages 1035–1044. Association for Com...

    2023

Pith tools

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