pith. sign in

arxiv: 2606.03138 · v1 · pith:DY74CORXnew · submitted 2026-06-02 · 💻 cs.IR

Section-Weighted Hybrid Approach for Legal Case Retrieval

Rajith Arulanandam , Nisansa de Silva This is my paper

Pith reviewed 2026-06-28 08:37 UTC · model grok-4.3

classification 💻 cs.IR
keywords legal case retrievalsection segmentationhybrid searchreciprocal rank fusionZ-score normalizationprecedent matchinginformation retrieval
0
0 comments X

The pith

A two-stage section-aware hybrid system retrieves analogous legal precedents more effectively than whole-document lexical or neural methods.

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

The paper presents a framework that segments legal judgments into facts, issues, decision, and reasoning sections using a deterministic LLM, then applies hybrid candidate retrieval followed by like-for-like section comparisons. It combines BM25 and dense vector searches via reciprocal rank fusion in stage one, then normalizes scores with query-wise Z-scores and aggregates them using learned section weights in stage two. A sympathetic reader would care because matching on specific legal elements rather than surface overlap can surface precedents that share reasoning structure, which matters for accurate legal analysis at scale.

Core claim

The paper claims that segmenting raw judgments offline into four sections, retrieving a high-recall candidate pool through parallel lexical and semantic search fused by RRF, and then performing fine-grained section-specific comparisons with Z-score normalization and learned weights produces consistent gains over strong baselines on a jurisdiction-scale benchmark while preserving high candidate coverage.

What carries the argument

Section-weighted aggregation of normalized lexical and semantic scores from like-for-like section comparisons, using query-wise Z-score normalization before applying learned weights.

If this is right

  • Top results can be returned with the matching section text, a grounded rationale, and party-stance labels.
  • The approach maintains high candidate coverage while improving ranking quality over pure lexical or neural baselines.
  • Query-wise normalization addresses the scale mismatch between lexical scores and cosine similarities before aggregation.
  • Fine-grained section comparisons enable matching on reasoning or facts independently rather than whole-document overlap.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same staged segmentation-plus-weighted-comparison pattern could apply to other structured document domains such as scientific papers or contracts.
  • If segmentation errors vary by jurisdiction, performance may degrade on legal systems with less standardized judgment formats.
  • Learned section weights could be made query-dependent to reflect different user intents such as fact-focused versus reasoning-focused searches.

Load-bearing premise

The deterministic LLM segmentation reliably and consistently identifies the sections across varied legal judgments without significant errors.

What would settle it

An experiment that replaces the LLM-derived section boundaries with random or fixed splits and measures whether the reported gains over baselines disappear.

read the original abstract

Finding truly analogous precedents requires capturing legal reasoning beyond surface word overlap. We present a two-stage, section-aware framework for legal case retrieval that first segments raw judgments into facts, issues, decision, and reasoning using a deterministic large language model (LLM) offline. In Stage 1, we combine parallel lexical (BM25) and semantic (dense ANN) whole-document searches via Reciprocal Rank Fusion (RRF) to form a high-recall candidate pool. In Stage 2, we perform fine-grained, like-for-like comparisons (e.g., query reasoning vs. candidate reasoning). To address the scale mismatch between unbounded lexical scores and cosine similarities, we apply query-wise Z-score normalization before aggregating signals with learned section weights. For the top results, the system returns the relevant section text with a concise, grounded rationale and party-stance labels. We evaluate on a jurisdiction-scale benchmark, demonstrating consistent gains over strong lexical and neural baselines while maintaining high candidate coverage

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript describes a two-stage section-aware hybrid retrieval framework for legal cases. Judgments are first segmented offline into facts/issues/decision/reasoning sections via a deterministic LLM. Stage 1 builds a high-recall candidate pool by fusing BM25 lexical and dense semantic retrieval with RRF. Stage 2 performs like-for-like section comparisons, applies query-wise Z-score normalization to reconcile score scales, and aggregates with learned section weights. The system returns top results with relevant section text, grounded rationales, and party-stance labels. It claims consistent gains over lexical and neural baselines on a jurisdiction-scale benchmark while preserving high candidate coverage.

Significance. If the reported gains are substantiated and the segmentation proves reliable, the work would advance legal IR by shifting emphasis from surface overlap to structured reasoning sections, with practical value in explainability. The Z-normalization step and learned weights address a common hybrid-retrieval scaling problem in a domain-appropriate way. The two-stage design (broad recall then fine-grained section matching) is a sensible response to the scale and structure of legal corpora.

major comments (2)
  1. [Abstract] Abstract: the central claim of 'consistent gains over strong lexical and neural baselines' is presented without any quantitative metrics, ablation results, dataset statistics, or error analysis, so the magnitude, statistical significance, and attribution of improvements to the section-weighted design cannot be evaluated.
  2. [Abstract] Abstract (pipeline description): the entire Stage-2 section-specific comparison and learned-weight aggregation presupposes that the deterministic LLM segmentation correctly partitions every judgment. No human agreement rates, segmentation error rates on held-out judgments, or sensitivity analysis are referenced; if segmentation error exceeds a few percent, mismatched sections invalidate the Z-normalized aggregation and the attribution of gains to the section-aware design.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments highlighting areas where the abstract could better support its claims. We address each point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'consistent gains over strong lexical and neural baselines' is presented without any quantitative metrics, ablation results, dataset statistics, or error analysis, so the magnitude, statistical significance, and attribution of improvements to the section-weighted design cannot be evaluated.

    Authors: We agree that the abstract would be strengthened by including concrete metrics. In the revised manuscript we will update the abstract to report key quantitative results from the evaluation section, including specific performance gains (e.g., nDCG@10 or MAP improvements over baselines), the scale of the jurisdiction benchmark, and a brief reference to ablation findings that attribute gains to the section-weighted stage. revision: yes

  2. Referee: [Abstract] Abstract (pipeline description): the entire Stage-2 section-specific comparison and learned-weight aggregation presupposes that the deterministic LLM segmentation correctly partitions every judgment. No human agreement rates, segmentation error rates on held-out judgments, or sensitivity analysis are referenced; if segmentation error exceeds a few percent, mismatched sections invalidate the Z-normalized aggregation and the attribution of gains to the section-aware design.

    Authors: The referee correctly notes the foundational role of segmentation quality. Our pipeline uses a fixed deterministic prompt, yet the manuscript provides no quantitative validation. We will add a new subsection describing the segmentation prompt, qualitative examples of output quality, and a limitations paragraph discussing the impact of potential segmentation mismatches. A full inter-annotator agreement study on held-out data is not present in the current work and would require new annotation effort. revision: partial

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper applies standard retrieval components (BM25, dense ANN, RRF, query-wise Z-score normalization, learned section weights) to an external jurisdiction-scale benchmark after offline LLM segmentation. No equations, predictions, or central claims reduce reported gains to fitted quantities by construction, nor rely on self-citation chains or imported uniqueness theorems. The segmentation step is an input assumption rather than a derived result, and the evaluation remains independent of the method's internal parameters.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The framework depends on the reliability of LLM-based section segmentation and the effectiveness of Z-score normalization for combining lexical and semantic scores; no free parameters or invented entities are explicitly introduced in the abstract.

free parameters (1)
  • learned section weights
    Weights are learned from data to aggregate section signals.

pith-pipeline@v0.9.1-grok · 5691 in / 1087 out tokens · 21499 ms · 2026-06-28T08:37:32.813074+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Legal Document Retrieval using Document Vector Embeddings and Deep Learning,

    K. Sugathadasa, B. Ayesha, N. de Silva, A. S. Perera, V . Jayawardana, D. Lakmal, and M. Perera, “Legal Document Retrieval using Document Vector Embeddings and Deep Learning,” inScience and information conference. Springer, 2018, pp. 160–175

    2018

  2. [2]

    The probabilistic relevance framework: BM25 and beyond,

    S. Robertson and H. Zaragoza, “The probabilistic relevance framework: BM25 and beyond,”Foundations and Trends in Information Retrieval, vol. 3, no. 4, pp. 333–389, 2009

    2009

  3. [3]

    Learning Sentence Embeddings in the Legal Domain with Low Resource Settings,

    S. Jayasinghe, L. Rambukkanage, A. Silva, N. de Silva, S. Perera, and M. Perera, “Learning Sentence Embeddings in the Legal Domain with Low Resource Settings,” inProceedings of the 36th Pacific Asia Conference on Language, Information and Computation, 2022, pp. 494– 502

    2022

  4. [4]

    Word Vector Embeddings and Domain Spe- cific Semantic based Semi-Supervised Ontology Instance Population,

    V . Jayawardana, D. Lakmal, N. de Silva, A. S. Perera, K. Sugathadasa, B. Ayesha, and M. Perera, “Word Vector Embeddings and Domain Spe- cific Semantic based Semi-Supervised Ontology Instance Population,” International Journal on Advances in ICT for Emerging Regions, vol. 10, no. 1, p. 1, 2017

    2017

  5. [5]

    Context sensitive verb similarity dataset for legal information extraction,

    G. Ratnayaka, N. de Silva, A. S. Perera, G. Kavirathne, T. Ariyarathna, and A. Wijesinghe, “Context sensitive verb similarity dataset for legal information extraction,”Data, vol. 7, no. 7, 2022

    2022

  6. [6]

    Learning interpretable legal case retrieval via knowledge-guided case reformulation,

    C. Deng, K. Mao, and Z. Dou, “Learning interpretable legal case retrieval via knowledge-guided case reformulation,” inEMNLP. ACL, 2024, pp. 1253–1265

    2024

  7. [7]

    CBR-RAG: Case-based reasoning for retrieval augmented generation in LLMs for legal question answering,

    N. Wiratunga, R. Abeyratne, L. Jayawardena, K. Martin, S. Massie, I. Nkisi-Orji, R. Weerasinghe, A. Liret, and B. Fleisch, “CBR-RAG: Case-based reasoning for retrieval augmented generation in LLMs for legal question answering,” inCase-Based Reasoning Research and Development: 32nd International Conference, ICCBR 2024, ser. Lecture Notes in Computer Scienc...

    2024

  8. [8]

    Explainable legal case matching via inverse optimal transport-based rationale extraction,

    W. Yu, Z. Sun, J. Xu, Z. Dong, X. Chen, H. Xu, and J.-R. Wen, “Explainable legal case matching via inverse optimal transport-based rationale extraction,” inProceedings of the 45th International ACM SI- GIR Conference on Research and Development in Information Retrieval (SIGIR ’22), 2022, pp. 657–668

    2022

  9. [9]

    Legalbench-RAG: A benchmark for retrieval-augmented generation in the legal domain,

    N. Pipitone and G. H. Alami, “Legalbench-RAG: A benchmark for retrieval-augmented generation in the legal domain,”arXiv preprint arXiv:2408.10343, 2024

    work page arXiv 2024

  10. [10]

    SCaLe-QA: Sri lankan case law embeddings for legal QA,

    L. Jayawardena, N. Wiratunga, R. Abeyratne, K. Martin, I. Nkisi-Orji, and R. Weerasinghe, “SCaLe-QA: Sri lankan case law embeddings for legal QA,” inProceedings of the SICSA Workshop on Real-World Applications of Large Language Models (REALLM 2024), ser. CEUR Workshop Proceedings, vol. 3822, 2024, pp. 47–55

    2024

  11. [11]

    ITALIAN-LEGAL-BERT models for improving natural language processing tasks in the italian legal domain,

    D. Licari and G. Comand `e, “ITALIAN-LEGAL-BERT models for improving natural language processing tasks in the italian legal domain,” Computer Law & Security Review, vol. 52, p. 105908, 2024

    2024

  12. [12]

    Hier-SPCNet: A legal statute hierarchy-based heterogeneous network for computing legal case document similarity,

    P. Bhattacharya, K. Ghosh, A. Pal, and S. Ghosh, “Hier-SPCNet: A legal statute hierarchy-based heterogeneous network for computing legal case document similarity,” inSIGIR, 2020, pp. 1657–1660

    2020

  13. [13]

    Law article-enhanced legal case matching: A causal learning approach,

    Z. Sun, J. Xu, X. Zhang, Z. Dong, and J.-R. Wen, “Law article-enhanced legal case matching: A causal learning approach,” inSIGIR, 2023, pp. 1549–1558

    2023

  14. [14]

    Logic rules as explanations for legal case retrieval,

    Z. Sun, K. Zhang, W. Yu, H. Wang, and J. Xu, “Logic rules as explanations for legal case retrieval,” inLREC-COLING, 2024

    2024

  15. [15]

    Learning fine-grained fact-article correspondence in legal cases,

    J. Ge, Y . Huang, X. Shen, C. Li, and W. Hu, “Learning fine-grained fact-article correspondence in legal cases,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 3694–3706, 2021

    2021

  16. [16]

    Precedent-enhanced legal judgment prediction with LLM and domain-model collaboration,

    Y . Wu, S. Zhou, Y . Liu, W. Lu, X. Liu, Y . Zhang, C. Sun, F. Wu, and K. Kuang, “Precedent-enhanced legal judgment prediction with LLM and domain-model collaboration,” inEMNLP. ACL, 2023, pp. 12 060– 12 075

    2023

  17. [17]

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. Xinget al., “Judging LLM-as-a-judge with MT-bench and chatbot arena,”arXiv preprint arXiv:2306.05685, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    An empirical study of LLM-as-a-judge for LLM evaluation: Fine-tuned judge model is not a general substitute for GPT-4,

    H. Huang, X. Bu, H. Zhou, Y . Qu, J. Liu, M. Yang, B. Xu, and T. Zhao, “An empirical study of LLM-as-a-judge for LLM evaluation: Fine-tuned judge model is not a general substitute for GPT-4,”arXiv preprint arXiv:2403.02839, 2024

    work page arXiv 2024

  19. [19]

    Humans or LLMs as the judge? a study on judgement bias,

    G. H. Chen, S. Chen, Z. Liu, F. Jiang, and B. Wang, “Humans or LLMs as the judge? a study on judgement bias,” inEMNLP, Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, Eds. Miami, Florida, USA: ACL, Nov. 2024, pp. 8301–8327

    2024

  20. [20]

    Synergistic Union of Word2Vec and Lexicon for Domain Specific Semantic Similarity,

    K. Sugathadasa, B. Ayesha, N. de Silva, A. S. Perera, V . Jayawardana, D. Lakmal, and M. Perera, “Synergistic Union of Word2Vec and Lexicon for Domain Specific Semantic Similarity,” in2017 IEEE International Conference on Industrial and Information Systems (ICIIS). IEEE, 2017, pp. 1–6

    2017

  21. [21]

    Legal Case Winning Party Prediction With Domain Specific Auxiliary Models,

    S. Jayasinghe, L. Rambukkanage, A. Silva, N. de Silva, and A. S. Perera, “Legal Case Winning Party Prediction With Domain Specific Auxiliary Models,” inProceedings of the 34th Conference on Computational Linguistics and Speech Processing (ROCLING 2022). ACL, 2022, pp. 205–213

    2022

  22. [22]

    Automatic Analysis of App Reviews Using LLMs,

    S. Gunathilaka and N. de Silva, “Automatic Analysis of App Reviews Using LLMs,” inProceedings of the Conference on Agents and Artificial Intelligence, 2025, pp. 828–839

    2025

  23. [23]

    Overview of the COLIEE 2025 competition: Legal case law and statute law information retrieval and entailment,

    R. Goebel, Y . Kano, M.-Y . Kim, C. Kwan, K. Satoh, H. Yamada, and M. Yoshioka, “Overview of the COLIEE 2025 competition: Legal case law and statute law information retrieval and entailment,” inProceedings of the 12th Competition on Legal Information Extraction and Entailment (COLIEE 2025) Workshop, Chicago, USA, 2025

    2025

  24. [24]

    SAILER: Structure-aware pre-trained language model for legal case retrieval,

    H. Li, Q. Ai, J. Chen, Q. Dong, Y . Wu, Y . Liu, C. Chen, and Q. Tian, “SAILER: Structure-aware pre-trained language model for legal case retrieval,” inSIGIR. ACM, 2023, pp. 1035–1044

    2023