Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Assessing the Performance Gap Between Lexical and Semantic Models for Information Retrieval With Formulaic Legal Language

T0 review · 4 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 In retrieval over formulaic CJEU legal language, BM25 beats off-the-shelf dense models on four of seven metrics, and its edge shows up when the query shares little verbatim wording with the target paragraph.

desk verdict A careful, reproducible extension of the CJEU paragraph retrieval benchmark; the headline claims hold up within the benchmark, but the citation-derived query construction is a load-bearing caveat the authors themselves acknowledge. read the letter →

arxiv 2506.12895 v1 pith:GWE6PLEB submitted 2025-06-15 cs.CL cs.IR

Larissa Mori , Carlos Sousa de Oliveira , Yuehwern Yih , Mario Ventresca This is my paper · ORCID
classification cs.CLcs.IR
keywords legalpassageretrievalBM25denseCJEUformulaiclanguagefine-tuninglexicaloverlapcitation-basedrelevance
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 sets out to establish when lexical, term-frequency-based retrieval and semantic, dense-embedding retrieval each do better on the formulaic language of Court of Justice of the European Union decisions. It shows that BM25, an unsupervised lexical scorer, is a strong baseline that beats off-the-shelf dense models on four of seven metrics in paragraph-level citation retrieval. Fine-tuning a dense model on domain data changes the balance, and the best fine-tuned model surpasses BM25 on most metrics. The paper's explanatory claim is that BM25's edge over dense models concentrates in cases where the query and the cited paragraph share little verbatim wording and where queries are longer, suggesting that dense models lose the relevant signal in surrounding context. If right, this tells legal retrieval system builders when a cheap lexical baseline suffices and when investing in domain fine-tuning is worth it.

What carries the argument

The central objects are BM25's probabilistic term-frequency score and the dense embeddings compared by cosine similarity. The mechanism that carries the argument is a set of overlap statistics between a query paragraph and the paragraph it cites: mean word-level edit distance, number of shared $N$-grams for $N=2,\dots,10$, and longest common subsequence. These statistics operationalize the paper's distinction between repetitive, quote-like legal language and more nuanced, low-overlap language, and they are what the statistical significance tests compare across success and failure cases.

What would settle it

Build a test set where professional legal researchers write realistic queries rather than extracting citations from judgments, have expert annotators mark relevant CJEU paragraphs, and rerun the seven metrics; if BM25's advantage over the zero-shot dense models shrinks or reverses on low-overlap, long queries, the paper's attribution of that advantage to lexical-overlap capture is wrong. A cheaper calculation on the existing data is to split long queries at the quoted span and measure whether dense models' recall rises toward BM25's, which would confirm or reject the extra-context-is-noise explanation.

Watch

Extended reading notes

Core claim

The authors establish, on a paragraph-level citation retrieval task built from CJEU decisions, that the statistical lexical model BM25 outperforms every off-the-shelf dense model on Recall@5, Recall@10, Recall@20 and nDCG@10, while the Ada-v2 embedding model narrowly beats it on Recall@1, MAP and MRR. Fine-tuning a small 33M-parameter BERT-based encoder on domain data flips the balance, and the larger LegalSBERT-ft model is best on all seven metrics. The paper's explanatory claim is that the performance gap is tied to how much the query repeats the cited text: when BM25 achieves perfect recall while a dense model gets zero, the query and the cited paragraph share significantly fewer $N$-grams, have larger mean word-level edit distance, and lower longest-common-subsequence scores than when both methods succeed; these queries are also longer on average. From this the authors conclude that lexical and dense models both handle repetitive, quote-like language, but BM25 is better at pulling the relevant passage out of longer, more nuanced queries where dense models are distracted by context beyond the quote.

Load-bearing premise

The comparison treats the paragraphs a CJEU judgment actually cites as the ground-truth relevant passages, even though the paper itself notes that citing paragraphs often quote or closely paraphrase the cited text, which is precisely the overlap BM25 exploits; if real legal research queries do not resemble such quotes, BM25's measured advantage may not carry over to practice.

Editorial extensions

If this is right

  • In resource-constrained legal retrieval settings, an unsupervised BM25 baseline can outperform large off-the-shelf embedding models on recall-oriented metrics, so organizations can defer dense retrieval until they have domain data and compute for fine-tuning.
  • Fine-tuning a 33M-parameter SBERT model on CJEU paragraphs beats zero-shot models four times its size, indicating that domain-specific contrastive tuning matters more than raw parameter count for this task.
  • Fine-tuning gains plateau: using 50% to 80% of the chronologically ordered data yields similar results, so collecting more training citations beyond a moderate amount brings diminishing returns.
  • The complementary strengths of BM25 and fine-tuned dense models make a BM25-first, dense-reranking pipeline a natural design, as the paper itself flags as future work.
  • On a broader 2015-2021 test set, the main ranking of methods is stable, so the conclusions are not an artifact of the default 2019-2021 test period.

Reading between the lines

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

  • Because the ground truth is citation-based, and CJEU citations often quote or closely paraphrase the cited paragraph, the benchmark likely overstates lexical models' practical advantage; a query set written by lawyers rather than extracted from judgments would test this.
  • The noise-from-context hypothesis is directly testable: for long queries where BM25 wins, truncate the query to its quoted span and check whether dense model recall rises toward BM25's; if it does not, the explanation would need revision.
  • The paper's overlap statistics could serve as a cheap formulaicity measure for other courts, predicting in advance whether lexical retrieval will suffice or whether fine-tuned semantic models are worth the cost.
  • If the noise hypothesis holds, dense models could be improved by training objectives or input segmentation that downweight non-quoted context, for instance by focusing the embedding on the quotation-bearing clause.
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

4 major / 5 minor

Summary. The paper evaluates lexical (BM25, TF-IDF) and dense retrieval models (zero-shot SBERT, SimCSE, Nomic, Ada-v2, Embedding-3-large; fine-tuned SBERT-ft and LegalSBERT-ft) on paragraph-level citation retrieval from CJEU decisions, using the dataset of Olsen et al. (2023). It reports standard IR metrics on a temporally split test set, concluding that BM25 is a strong baseline that outperforms off-the-shelf dense models on 4 of 7 metrics, that fine-tuned dense models surpass BM25 on most metrics, and that when BM25 outperforms dense models the query and cited passage have lower lexical overlap and queries tend to be longer. The paper also presents a fine-tuning data-amount ablation and discusses temporal robustness.

Significance. If the findings hold, the paper provides a useful benchmark contribution: it extends the original CJEU paragraph-retrieval evaluation from 3 methods and 1 metric to 10 methods and 7 metrics, releases code and preprocessed data, and offers a concrete, reproducible analysis of when lexical matching is competitive with dense retrieval in legal text. The qualitative overlap analysis, with examples, is a useful starting point for understanding model failure modes. However, the benchmark's citation-derived queries systematically contain verbatim quotes or close paraphrases of the target paragraph, as the paper itself acknowledges in Section 8. This is not an internal inconsistency, but it is a load-bearing validity limitation for the headline claim about BM25's strength relative to dense models in 'nuanced' scenarios.

major comments (4)
  1. [Section 8 / Table 4] The benchmark validity limitation is load-bearing rather than a minor caveat. Queries are citing paragraphs, and as Section 8 states, citing paragraphs frequently contain direct quotes or paraphrases of the cited text, which rewards lexical matching methods such as BM25. The central claim that BM25 outperforms off-the-shelf dense models in 4 of 7 metrics is measured only on this citation-derived benchmark. To support the broader claim that BM25 is superior in 'more nuanced scenarios,' the authors should add an evaluation that does not inherit this lexical-overlap bias, for example queries written by legal practitioners, or a controlled paraphrase/synonym substitution of queries that removes verbatim overlap. At minimum, reporting performance on the subset of test queries with low lexical overlap (e.g., low LCS) would separate the effect of benchmark construction from the model comparison.
  2. [Section 4 / Table 4] All results in Table 4 are from a single run and are reported without confidence intervals, bootstrap intervals, or significance tests. Several differences that support the '4 out of 7 metrics' claim are small, for example BM25 versus Ada-v2 on MAP (0.5132 vs 0.5150) and on nDCG@10 (0.5767 vs 0.5696). Without uncertainty quantification, the relative ranking of BM25 and the best zero-shot dense model is not robustly established. The authors should provide variance estimates over multiple seeds (for the fine-tuned models) or a paired bootstrap over test queries for all methods.
  3. [Section 5.2 / Figures 2–3] The t-tests in Section 5.2 compare groups defined by extreme retrieval outcomes: queries where BM25 has 100% recall and the dense model has 0% recall versus queries where both have 100% recall. This is a selection-on-outcome design, so the observed differences in lexical overlap, edit distance, and LCS are descriptive correlations, not evidence that low overlap causes BM25 to outperform dense models. The causal story that 'additional information adds noise' in dense models is supported by a single hand-picked example (Example 4 in Section 5.1) and no controlled manipulation. A more direct test would hold the quoted portion of the query fixed and vary the amount of extra context, then measure dense-model and BM25 performance as a function of that manipulation.
  4. [Section 6 / Table 6] The fine-tuning ablation in Section 6 reports single-run results for each data percentage. The claim that performance plateaus is based on small absolute differences, and the non-monotonicity (e.g., Recall@1 is 42.13% at 50% data but 41.80% at 80% data) could easily be training noise. The authors should report multiple fine-tuning runs per percentage and show stability, or temper the plateau claim accordingly.
minor comments (5)
  1. [Equation (1)] The BM25 formula is garbled in the rendered text, with placeholder symbols (e.g., 'Q1' and 'R' in the narrative but inconsistent notation in the equation). Please provide a clean, correctly typeset equation with consistent notation for k1 and b.
  2. [Section 5.1] The qualitative analysis samples only 3 examples per scenario without stating the random seed. Given the small sample and the illustrative intent, please state that the examples are illustrative and avoid drawing general conclusions from them alone.
  3. [Appendix G] The paper analyzes the case where BM25 succeeds and dense models fail, but the complementary case where BM25 has 0% recall and Ada-v2 has 100% recall is only briefly described in Appendix G and is not quantitatively compared using the same overlap metrics. Adding that comparison would strengthen the analysis and help separate benchmark bias from genuine model differences.
  4. [Throughout] There are several typos and OCR artifacts, for example 'the the goal' in Section 2.1, 'peformance' in Section 8, 'Ada-vs' in Table 14, and '!ne-tuning' throughout. A careful proofreading pass is needed.
  5. [Tables 3 and 8] Tables 3 and 8 are duplicated content. Please remove the duplicate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the model comparison is a direct empirical evaluation on an external citation-labeled dataset; the acknowledged lexical-overlap bias is a validity limitation, not a circular derivation.

full rationale

The paper makes no fitted-parameter-as-prediction move, imports no load-bearing result from the authors' own prior work, and does not define its key quantities in terms of each other. BM25 and dense model scores are computed on a test set split temporally from the external CJEU dataset of [16]; fine-tuned models are trained on earlier citations and evaluated on later ones. The headline comparison (Table 4) is a measured ranking, not a derivation from assumptions. The qualitative/quantitative gap analysis (Section 5) compares overlap metrics between groups defined by model outcomes; this is a post-hoc description, not a prediction forced by construction. Section 8 explicitly concedes that citation labels, which favor lexical matching, may overestimate BM25 relative to practical legal research: 'using citations as the measure of relevance tends to favor models that are particularly effective at matching lexical patterns between queries and citations.' That is a validity limitation on generalization, not an internal circularity. No self-citation chain is load-bearing; the preliminary-version footnote is not used to justify any claim. Score 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper's claims rest on the citation-as-relevance assumption and the temporal data split, not on any new theoretical construct. No new entities are introduced.

free parameters (2)
  • TF-IDF vocabulary size (top-K) = 5,000
    Chosen by hand for the TF-IDF baseline in Section 3.2.1; not fitted to the target task and not central to the main claims.
  • TF-IDF n-gram range = 1 to 2
    Fixed by hand in Section 3.2.1; a design choice for the baseline, not a fitted parameter.
assumptions (3)
  • domain assumption Citations between CJEU paragraphs are valid ground-truth relevance labels
    Motivates the whole dataset and evaluation; the paper acknowledges in Section 8 that this may favor lexical methods.
  • domain assumption Temporal split (train 1979-2016, validation 2017-2018, test 2019-2021) approximates real-world retrieval from past precedent
    Used in Section 3.1 to define training and test; appendix H checks an alternative 2015-2021 test window.
  • domain assumption The preprocessed dataset is representative of CJEU paragraph-level citations after removing 269 empty paragraphs and 4,851 French paragraphs
    Data cleaning in Section 3.1 could affect statistics, but the removals are documented.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Assessing the Performance Gap Between Lexical and Semantic Models for Information Retrieval With Formulaic Legal Language." pith.science (2026). https://pith.science/paper/GWE6PLEB

@misc{pith2026250612895,
  author       = {Pith},
  title        = {Pith review of: Assessing the Performance Gap Between Lexical and Semantic Models for Information Retrieval With Formulaic Legal Language},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GWE6PLEB}},
  note         = {Machine review of arXiv:2506.12895}
}
read the original abstract

Legal passage retrieval is an important task that assists legal practitioners in the time-intensive process of finding relevant precedents to support legal arguments. This study investigates the task of retrieving legal passages or paragraphs from decisions of the Court of Justice of the European Union (CJEU), whose language is highly structured and formulaic, leading to repetitive patterns. Understanding when lexical or semantic models are more effective at handling the repetitive nature of legal language is key to developing retrieval systems that are more accurate, efficient, and transparent for specific legal domains. To this end, we explore when this routinized legal language is better suited for retrieval using methods that rely on lexical and statistical features, such as BM25, or dense retrieval models trained to capture semantic and contextual information. A qualitative and quantitative analysis with three complementary metrics shows that both lexical and dense models perform well in scenarios with more repetitive usage of language, whereas BM25 performs better than the dense models in more nuanced scenarios where repetition and verbatim~quotes are less prevalent and in longer queries. Our experiments also show that BM25 is a strong baseline, surpassing off-the-shelf dense models in 4 out of 7 performance metrics. However, fine-tuning a dense model on domain-specific data led to improved performance, surpassing BM25 in most metrics, and we analyze the effect of the amount of data used in fine-tuning on the model's performance and temporal robustness. The code, dataset and appendix related to this work are available on: https://github.com/larimo/lexsem-legal-ir.

Figures

Figures reproduced from arXiv: 2506.12895 by the authors.

Figure 1
Figure 1. Example of a citing paragraph (top) and a cited [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Average number of 𝑁-grams in common between query and cited paragraph when both BM25 and Ada-v2 have perfect Recall@5 (Both perfect recall), and when BM25 has 100% recall and Ada-v2 has 0% Recall@5 (BM25 > Ada-v2), for 𝑁 = 2, 3, ..., 10. The shaded regions indicate the standard deviation. edits are necessary to turn the query into the cited document in the !rst scenario compared to 66.81 (std. 39.84) in the second. … view at source ↗
Figure 3
Figure 3. The three metrics analyzed point to the same conclusion that there is a lower presence of verbatim quotes and paraphrases be￾tween the query and cited documents in the scenario in which BM25 performs well and Ada-v2 does not. Thus, both lexical and dense models perform well in the scenarios with more repetitive usage of language, whereas BM25 performs better than the dense models in more nuanced scenarios where repe… view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Comparison between Mean Edit Distance and LCS [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]
Figure 5
Figure 5. Figure 5: Average number of 𝑁-grams in common between query and cited paragraph when both BM25 and SBERT-ft have perfect Recall@20 (Both perfect recall), and when BM25 has 100% recall and SBERT-ft has 0% Recall@20 (BM25 > SBERT-!), for 𝑁 = 2, 3, ..., 10. The shaded regions indic…
Figure 6
Figure 6. Figure 6: Comparison between Mean Edit Distance and LCS [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 12 canonical work pages

  1. [1]

    Loıc Azoulai. 2011. The retained powers’ formula in the case law of the european court of justice: EU law as total law. Eur. J. Legal Stud. 4 (2011), 178. Publisher: HeinOnline

    work page 2011

  2. [2]

    Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, and Ion Androutsopoulos. 2020. LEGAL-BERT: The Muppets straight out of Law School. In Findings of the Association for Computational Linguistics: EMNLP 2020 . Association for Computational Linguistics, Online, 2898–2904. doi:10.18653/v1/ 2020.!ndings-emnlp.261

    doi:10.18653/v1/ 2020

  3. [3]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. doi:10.48550/arXiv.1810.04805 arXiv:1810.04805 [cs]

    arXiv doi:10.48550/arxiv.1810.04805 2019

  4. [4]

    Yi Feng, Chuanyi Li, and Vincent Ng. 2024. Legal Case Retrieval: A Survey of the State of the Art. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 6472–6485. https://aclant...

    work page 2024

  5. [5]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple Contrastive Learning of Sentence Embeddings. In Proceedings of the 2021 Conference on Em- pirical Methods in Natural Language Processing , Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih (Eds.). Association for Compu- tational Linguistics, Online and Punta Cana, Domi...

    doi:10.18653/v1/2021.emnlp-main.552 2021

  6. [6]

    Randy Goebel, Yoshinobu Kano, Mi-Young Kim, Juliano Rabelo, Ken Satoh, and Masaharu Yoshioka. 2023. Summary of the competition on legal information, ex- traction/entailment (COLIEE) 2023. In Proceedings of the nineteenth international conference on arti!cial intelligence and law . 472–480

    work page 2023

  7. [7]

    Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun-hsuan Sung, Laszlo Lukacs, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. 2017. E#cient Natural Language Response Suggestion for Smart Reply. http://arxiv.org/abs/1705.00652 arXiv:1705.00652 [cs]

    arXiv 2017

  8. [8]

    Abe Bohan Hou, Orion Weller, Guanghui Qin, Eugene Yang, Dawn Lawrie, Nils Holzenberger, Andrew Blair-Stanek, and Benjamin Van Durme. 2024. CLERC: A 9For our computations, we use a NVIDIA Geforce RTX 3050 Ti. Assessing the Performance Gap Between Lexical and Semantic Models for Information Retrieval With Formulaic Legal Language ICAIL ’25, 2025, Chicago, I...

    arXiv 2024

Show all 26 references
  1. [9]

    Je" Johnson, Matthijs Douze, and Hervé Jégou. 2017. Billion-scale similarity search with GPUs. http://arxiv.org/abs/1702.08734 arXiv:1702.08734 [cs]

    2017 arXiv

  2. [10]

    Vladimir Karpukhin, Barlas O%uz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. http://arxiv.org/abs/2004.04906 arXiv:2004.04906 [cs]

    2020 arXiv

  3. [11]

    Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent re- trieval for weakly supervised open domain question answering. arXiv preprint arXiv:1906.00300 (2019)

    2019 arXiv

  4. [12]

    Daniel Locke and Guido Zuccon. 2022. Case law retrieval: problems, methods, challenges and evaluations in the last 20 years. http://arxiv.org/abs/2202.07209 arXiv:2202.07209 [cs]

    2022 arXiv

  5. [13]

    Robert Mahari, Dominik Stammbach, Elliott Ash, and Alex ‘Sandy’ Pentland

  6. [14]

    Enys Mones, Piotr Sapie&y’ski, Simon Thordal, Henrik Palmer Olsen, and Sune Lehmann. 2021. Emergence of network e"ects and predictability in the judicial system. Scienti!c Reports 11, 1 (Feb. 2021), 2740. doi:10.1038/s41598-021-82430-x

    2021 doi

  7. [15]

    Morris, Brandon Duderstadt, and Andriy Mulyar

    Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. 2024. Nomic embed: Training a reproducible long context text embedder. arXiv: 2402.01613 [cs.CL]

    2024 arXiv

  8. [16]

    Henrik Palmer Olsen, Nicolas Garneau, Yannis Panagis, Johan Lindholm, and Anders Søgaard. 2023. Re-Framing Case Law Citation Prediction from a Paragraph Perspective. In Frontiers in Arti!cial Intelligence and Applications, Giovanni Sileno, Jerry Spanakis, and Gijs Van Dijck (E...

    2023 doi

  9. [17]

    Yannis Panagis, Ur(ka )adl, and Fabien Tarissan. 2017. Giving every case its (legal) due-the contribution of citation networks and text similarity techniques to legal studies of European Union law. In Legal knowledge and information systems . IOS Press, 59–68

    2017

  10. [18]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. http://arxiv.org/abs/1908.10084 arXiv:1908.10084 [cs]

    2019 arXiv

  11. [19]

    Stephen Robertson, Hugo Zaragoza, and others. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends® in Information Retrieval 3, 4 (2009), 333–389

    2009

  12. [20]

    Guilherme Moraes Rosa, Ruan Chaves Rodrigues, Roberto Lotufo, and Rodrigo Nogueira. 2021. Yes, BM25 is a Strong Baseline for Legal Case Retrieval. http: //arxiv.org/abs/2105.05686 arXiv:2105.05686 [cs]

    2021 arXiv

  13. [21]

    Li Tang and Simon Clematide. 2021. Searching for Legal Documents at Paragraph Level: Automating Label Generation and Use of an Extended Attention Mask for Boosting Neural Models of Semantic Similarity. In Proceedings of the Natural Legal Language Processing Workshop 2021. Asso...

    2021 doi

  14. [22]

    Aleksandar Trklja. 2017. A corpus investigation of formulaicity and hybridity in legal language: a case of EU case law texts. In Phraseology in legal and institutional settings. Routledge, 89–108

    2017

  15. [23]

    Santosh T.y.s.s., Rashid Haddad, and Matthias Grabmair. 2024. ECtHR-PCR: A Dataset for Precedent Understanding and Prior Case Retrieval in the European Court of Human Rights. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resou...

    2024

  16. [26]

    ecting the !nancial interests of the European Union in customs mat- ters. It added, in the same point, that it is for the national court to give full e

    Ada-v2’s and Emb-3-large’s counts of parameters are not made public by OpenAI. Also, a count of parameters and embedding dimension does not apply to BM25 and TD-IDF, which are thus omitted. Method # params. embed. dim. SBERT/SBERT-ft 33M 384 LegalSBERT-ft 110M 768 SimCSE 125M ...

    2016

  17. [2023]

    http://arxiv

    LePaRD: A Large-Scale Dataset of Judges Citing Precedents. http://arxiv. org/abs/2311.09356 arXiv:2311.09356 [cs]

    arXiv

  18. [5483]

    Each deci- sion that has at least one citation has an average of 13.09 outbound citations, with standard deviation of 13.64

    https://aclanthology.org/2024.lrec-main.486 A Further dataset statistics We provide further dataset statistics at the decision level. Each deci- sion that has at least one citation has an average of 13.09 outbound citations, with standard deviation of 13.64. Each decision that...

    2024

Pith tools

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