Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Generative Chinese statute retrieval with structured docids and multi-task training beats sparse, dense, and legal baselines on real lay queries.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 06:59 UTC pith:U7QYPXFR

load-bearing objection Solid applied GR for Chinese statutes on STARD; gains are real but mostly from multi-task data (esp. LLM pseudo-queries), not the fancy multi-field docid. the 3 major comments →

arxiv 2607.11109 v1 pith:U7QYPXFR submitted 2026-07-13 cs.IR cs.CL

Generative Chinese Statute Retrieval

Yiteng Tu , Zitao Su , Weihang Su , Xuanyi Chen , Yueyue Wu , Yiqun Liu , Min Zhang , Qingyao Ai This is my paper
classification cs.IR cs.CL
keywords Statute RetrievalGenerative RetrievalLegalStructured DocidChinese LawMulti-task TrainingPseudo-query Augmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Everyday legal questions are colloquial and narrative, while statutes are formal, hierarchical, and technical, so ordinary retrieval often fails to connect a lay description of a problem to the right articles. This paper claims that treating statute retrieval as sequence generation—having a model decode multi-field structured identifiers that encode legal branch, provision type, hierarchy, objects, title, and a short summary—lets the model internalize the statute corpus and map informal queries directly to the right articles. Training combines three tasks: reconstructing each article’s docid from its full text, aligning LLM-generated pseudo-queries that mimic non-professional language, and supervised fine-tuning on real query–statute pairs. On the STARD benchmark of 1,543 real public consultations over 55,348 Chinese statutes, the resulting system (GCSR) consistently outperforms strong sparse, dense, and legal-domain retrievers, including a fine-tuned dense baseline. A sympathetic reader would care because statute retrieval is the entry point for legal advice, drafting, and retrieval-augmented legal systems; a method that reliably surfaces the right provisions from plain-language questions improves access to law for both professionals and the public.

Core claim

GCSR shows that generative retrieval—decoding multi-granularity structured document identifiers that encode Chinese legal hierarchy and semantics, trained with statute indexing, LLM pseudo-query augmentation, and supervised retrieval—consistently outperforms strong sparse, dense, and legal-domain baselines on the STARD Chinese statute retrieval benchmark of real non-professional queries.

What carries the argument

Multi-granularity structured docid: a concatenated, field-marked sequence (legal branch, sub-category, provision type, legislation level, involved objects, abbreviated title, content summary) that serves as the generation target and supplies a hierarchical semantic path from colloquial query to specific article.

Load-bearing premise

The claim rests on the idea that rule- and LLM-built multi-field docids plus ten synthetic colloquial pseudo-queries per article create a target space and training distribution that truly bridges real lay queries to formal statutes.

What would settle it

On a held-out set of real public legal consultations with no training leakage, measure whether GCSR’s Hits@3 remains significantly above a strong fine-tuned dense retriever after removing the pseudo-query task or after replacing the multi-field structured docid with atomic or title-only identifiers; a collapse or loss of significance would falsify the central claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Statute retrieval can be cast as end-to-end docid generation rather than index-then-retrieve, with the model parameters serving as the corpus index.
  • Structured hierarchical identifiers reduce ambiguity among similar articles and guide decoding better than atomic IDs, titles alone, or product-quantization codes.
  • Pseudo-query augmentation that mimics non-professional language is necessary to close the colloquial–legal gap; indexing alone is not enough.
  • Accurate generative statute retrieval can serve as a foundation for legal question answering, compliance checking, and retrieval-augmented legal generation.
  • The same structured-docid and multi-task pattern is intended to extend to other civil-law jurisdictions and languages with hierarchical codes.

Where Pith is reading between the lines

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

  • If the multi-field path is doing the real work, simpler hierarchical prefixes (branch + legislation + article number) might capture most of the gain without full LLM summarization.
  • The large drop when pseudo-query or supervised tasks are removed suggests that generative indexing mainly succeeds when the training distribution already contains colloquial-to-formal pairs, not from pure corpus memorization.
  • Failure modes on rare or newly enacted statutes would reveal how brittle internalization is when the corpus is no longer static.
  • The same design could be stress-tested for cross-reference-heavy retrieval, where one article’s applicability depends on another’s hierarchy.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes GCSR, a generative retrieval framework for Chinese statute retrieval that casts retrieval as autoregressive generation of multi-granularity structured docids (B⊕C⊕T⊕L⊕O⊕A⊕S) encoding legal branch, subcategory, provision type, legislation level, involved objects, abbreviated title, and an LLM-generated summary. Training combines three tasks under a shared NLL objective (Eq. 2): statute indexing (article text → docid), pseudo-query augmentation (10 Qwen2.5-14B queries per article), and supervised fine-tuning on STARD query–statute pairs. On STARD (1,543 queries; 55,348 articles), mT5-base GCSR reports Hits@3/5/10/20 of 0.522/0.536/0.544/0.547, outperforming sparse (BM25, QL), zero-shot dense (BGE, RoBERTa), legal pre-trained (SAILER, Lawformer), and a STARD-fine-tuned Dense baseline (0.468 Hits@3), with pairwise significance marks. Ablations (Tables 3–4) compare docid variants and task ablations.

Significance. If the gains are attributable to generative indexing with structured docids rather than primarily to extra synthetic data, the work would be a useful contribution to legal IR: statute corpora are relatively static and hierarchical, so model-based indexing is a natural fit, and STARD is a realistic non-professional consultation benchmark. The paper is clear about the colloquial–formal gap, provides concrete docid instantiations (Table 1), and reports standard Hits@K with significance tests. Strengths include a full multi-task recipe, constrained Trie decoding, and ablations that at least partially separate docid design from training tasks. The result is of practical interest for Chinese legal search and RAG grounding, though currently limited to one language and one benchmark.

major comments (3)
  1. [Table 2, §3.2, Table 4] Table 2 vs. §3.2 / Table 4: The central claim that generative retrieval with multi-granularity docids is what drives superiority over dual-encoder baselines is not isolated. Dense is fine-tuned only on STARD annotated pairs (official setting). GCSR additionally receives full-corpus Statute Indexing plus ~550k Qwen2.5-14B pseudo-query pairs before SRT. Table 4 shows removing PAT alone drops Hits@3 from 0.522 to 0.387 (and removing SRT to 0.373), so most of the lift tracks data augmentation. Without a Dense (or other dual-encoder) control trained on the same pseudo-queries and indexing-style positives, the paper cannot attribute the 0.522 vs 0.468 gap to the generative paradigm rather than training distribution.
  2. [Table 3, §3.1] Table 3: Multi-granularity structure is presented as a core contribution, yet Atomic (random integer) docids already reach Hits@3 0.515 and Semantic 0.517 versus GCSR 0.522; Title and PQ lag more clearly. The hierarchical fields (B,C,T,L,O) therefore appear to contribute little once the multi-task training distribution is fixed. Either strengthen the case for structure (e.g., collision rates, error analysis by hierarchy level, or a controlled comparison holding training data fixed) or temper claims that multi-granularity encoding is the primary driver of gains.
  3. [§4.1–4.3, Conclusions] §4.1–4.3: Evaluation is confined to a single Chinese benchmark (STARD) with no multi-run variance, no second corpus/language, and no analysis of invalid or near-miss docids under Trie-constrained beam search. Given that the abstract and conclusion advertise broader legal IR and downstream reasoning potential, at least one additional evaluation setting, or explicit reporting of run variance and failure modes, is needed to support the generality of the effectiveness claim.
minor comments (5)
  1. [Headers / title page] Placeholder venue line "Conference acronym 'XX, June 03–05, 2018, Woodstock, NY" remains throughout headers and should be removed or replaced.
  2. [§3.1] Docid field construction (rule-based classifiers for T/L/O, keyword mapping for B/C, DeepSeek summary for S) is only sketched; a short appendix on rules, collision rates, and summary quality would improve reproducibility.
  3. [Table 3] Table 3 Hits@20: Semantic (0.548) slightly exceeds GCSR (0.547); note whether this is within noise and whether significance was tested among docid variants.
  4. [Eq. (1)] Eq. (1) uses ⊕ for string concatenation; state explicitly that fields are serialized with delimiters/special tokens so the decoder can learn field boundaries.
  5. [§2] Related work on generative retrieval is adequate but could briefly situate against other legal statute-retrieval graph/GNN lines (e.g., Louis et al.) beyond the STARD citation.

Circularity Check

0 steps flagged

No significant circularity: GCSR claims rest on empirical Hits@K comparisons after multi-task training, not on any derivation that reduces to its inputs by construction.

full rationale

The paper reformulates statute retrieval as seq2seq generation of multi-granularity structured docids (B⊕C⊕T⊕L⊕O⊕A⊕S), trains mT5 via three tasks (statute indexing on full articles, 10 Qwen2.5-14B pseudo-queries per article, and supervised fine-tuning on STARD annotated pairs), then evaluates Hits@{3,5,10,20} against sparse, dense, legal, and fine-tuned baselines on the STARD corpus of 55k articles and 1.5k queries. Ablations (Tables 3–4) isolate docid variants and task ablations as ordinary empirical controls. There is no equation, uniqueness theorem, or fitted constant that is later re-presented as a prediction; performance numbers are produced by beam-search decoding constrained by a Trie and compared via t-tests. Self-citations (STARD [22], SAILER [7], Lawformer [33], etc.) supply the public benchmark and related legal IR context by overlapping authors, but the central claim—that the generative setup outperforms the official STARD Dense baseline—is an independent experimental outcome, not forced by those citations or by definitional identity. Standard supervised use of a dataset’s training split is not circular when metrics are reported on the evaluation set. No self-definitional loop, fitted-input-as-prediction, or ansatz smuggled via self-citation appears in the derivation chain.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central performance claim rests on standard seq2seq training assumptions, a hand-designed multi-field docid built with rules and external LLMs, fixed training hyperparameters, and the STARD query–statute distribution. No physical constants; free parameters are ML hyperparameters and design choices that affect the reported Hits@K.

free parameters (5)
  • learning_rate = 1e-3
    Set to 1e-3 for 300k-step mT5-base training; not derived, chosen for optimization.
  • training_steps = 300000
    300k steps on 4 A100s; schedule choice that can affect final Hits@K.
  • pseudo_queries_per_article = 10
    Fixed at 10 Qwen2.5-14B generations per statute; ablation shows this task is load-bearing.
  • beam_size = 20
    Beam search size 20 at inference for top candidates; directly tied to Hits@20 evaluation.
  • per_device_batch_size = 48
    Batch size 48 per device; training dynamics hyperparameter.
axioms (5)
  • ad hoc to paper Chinese statutes admit a stable multi-field discrete encoding (branch, subcategory, provision type, legislation level, objects, abbreviated title, summary) that is learnable as an autoregressive target.
    Introduced in §3.1 as the multi-granularity structured docid; fields mix taxonomy, rules, and LLM summary.
  • domain assumption A relatively static, hierarchical statute corpus can be internalized in generative model parameters more effectively for retrieval than dual-encoder similarity for this task.
    Stated in Introduction and Related Work as motivation for GR over retrieve-then-rerank.
  • domain assumption LLM-generated informal pseudo-queries approximate non-professional legal consultation language well enough to train the colloquial–formal mapping.
    §3.2 Pseudo-query Augmentation Task; ablation Table 4 shows large drop without PAT.
  • standard math Negative log-likelihood on gold docid tokens (Eq. 2) is a valid end-to-end objective for retrieval quality measured by Hits@K.
    Standard autoregressive training used for all three tasks in §3.2.
  • domain assumption Trie-constrained decoding guarantees only valid corpus docids, so generation errors reduce to ranking among existing articles.
    §4.2 evaluation protocol following prior GR work.
invented entities (2)
  • Multi-granularity structured docid (B⊕C⊕T⊕L⊕O⊕A⊕S) no independent evidence
    purpose: Provide a hierarchical, semantic, collision-resistant generation target for each Chinese statutory article.
    Defined in §3.1 and Table 1; constructed via taxonomy mapping, rules, abbreviation, and DeepSeek summary—not an independently measured natural object.
  • GCSR multi-task training regime (SIT + PAT + SRT) no independent evidence
    purpose: Internalize statute text and align colloquial queries to structured docids in one generative model.
    §3.2 defines the three tasks and unified NLL objective; existence is methodological, not an external entity.

pith-pipeline@v1.1.0-grok45 · 14412 in / 3521 out tokens · 57196 ms · 2026-07-14T06:59:58.893428+00:00 · methodology

0 comments
read the original abstract

Statute retrieval is a fundamental task in legal information retrieval, yet existing approaches struggle to bridge the gap between colloquial legal queries and formal statutory language. In this paper, we propose GCSR, a generative statute retrieval framework that reformulates statute retrieval as a sequence generation problem and internalizes statutory knowledge into a generative model. Specifically, we propose a multi-granularity structured docid that encodes legal hierarchy and semantic information, together with a multi-task training strategy. Experiments show that GCSR consistently outperforms strong sparse, dense, and legal-domain baselines. Our results demonstrate the effectiveness of generative retrieval for statute retrieval and highlight its potential for broader legal information access and downstream legal reasoning tasks.

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

44 extracted references · 10 linked inside Pith

  1. [1]

    Francesco Borgesano, Annarita De Maio, Pasquale Laghi, and Roberto Musmanno

  2. [2]

    Artificial intelligence and justice: a systematic literature review and future research perspectives on Justice 5.0.European Journal of Innovation Management 28, 11 (2025), 349–385

    2025

  3. [3]

    2023.Chinese law: Towards an understanding of Chinese law, its nature and developments

    Jianfu Chen. 2023.Chinese law: Towards an understanding of Chinese law, its nature and developments. Vol. 3. Martinus Nijhoff Publishers

    2023

  4. [4]

    Szu-Ju Chen, Jing Jin, Sheng-Lun Wei, Chien-Hung Chen, and Hsin-Hsi Chen

  5. [5]

    InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Retrieving the Right Law: Enhancing Legal Search with Style Translation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2951–2955

  6. [6]

    Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, and Ziqing Yang. 2021. Pre- training with whole word masking for chinese bert.IEEE/ACM Transactions on Audio, Speech, and Language Processing29 (2021), 3504–3514

    2021

  7. [7]

    Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2020. Au- toregressive entity retrieval.arXiv preprint arXiv:2010.00904(2020)

    Pith/arXiv arXiv 2020

  8. [8]

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

    Pith/arXiv arXiv 2025

  9. [9]

    Haitao Li, Qingyao Ai, Jia Chen, Qian Dong, Yueyue Wu, Yiqun Liu, Chong Chen, and Qi Tian. 2023. SAILER: structure-aware pre-trained language model for legal case retrieval. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1035–1044

    2023

  10. [10]

    Haitao Li, Weihang Su, Changyue Wang, Yueyue Wu, Qingyao Ai, and Yiqun Liu

  11. [11]

    Thuir@ coliee 2023: Incorporating structural knowledge into pre-trained language models for legal case retrieval.arXiv preprint arXiv:2305.06812(2023)

    Pith/arXiv arXiv 2023

  12. [12]

    Daniel Locke and Guido Zuccon. 2022. Case law retrieval: problems, methods, challenges and evaluations in the last 20 years.arXiv preprint arXiv:2202.07209 (2022)

    Pith/arXiv arXiv 2022

  13. [13]

    Antoine Louis, Gijs Van Dijck, and Gerasimos Spanakis. 2023. Finding the law: Enhancing statutory article retrieval via graph neural networks.arXiv preprint arXiv:2301.12847(2023)

    Pith/arXiv arXiv 2023

  14. [14]

    Yixiao Ma, Yueyue Wu, Weihang Su, Qingyao Ai, and Yiqun Liu. 2023. CaseEn- coder: A Knowledge-enhanced Pre-trained Model for Legal Case Encoding.arXiv preprint arXiv:2305.05393(2023)

    Pith/arXiv arXiv 2023

  15. [15]

    2023.The civil law tradition: an introduction to the legal systems of Europe and Latin America

    John Merryman and Rogelio Pérez-Perdomo. 2023.The civil law tradition: an introduction to the legal systems of Europe and Latin America. Stanford University Press

    2023

  16. [16]

    Donald Metzler, Yi Tay, Dara Bahri, and Marc Najork. 2021. Rethinking search: making domain experts out of dilettantes. InAcm sigir forum, Vol. 55. ACM New York, NY, USA, 1–27

    2021

  17. [17]

    Ha-Thanh Nguyen, Manh-Kien Phi, Xuan-Bach Ngo, Vu Tran, Le-Minh Nguyen, and Minh-Phuong Tu. 2024. Attentive deep neural networks for legal document retrieval.Artificial Intelligence and Law32, 1 (2024), 57–86

    2024

  18. [18]

    Jay M Ponte and W Bruce Croft. 2017. A language modeling approach to in- formation retrieval. InACM SIGIR Forum, Vol. 51. ACM New York, NY, USA, 202–208

    2017

  19. [19]

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

    2009

  20. [20]

    Carlo Sansone and Giancarlo Sperlí. 2022. Legal information retrieval systems: state-of-the-art and open issues.Information Systems106 (2022), 101967

    2022

  21. [21]

    Jaromír Šavelka and Kevin D Ashley. 2022. Legal information retrieval for understanding statutory terms.Artificial Intelligence and Law30, 2 (2022), 245– 289

    2022

  22. [22]

    Yunqiu Shao, Haitao Li, Yueyue Wu, Yiqun Liu, Qingyao Ai, Jiaxin Mao, Yixiao Ma, and Shaoping Ma. 2023. An intent taxonomy of legal case retrieval.ACM Transactions on Information Systems42, 2 (2023), 1–27

    2023

  23. [23]

    Weihang Su, Qingyao Ai, Yueyue Wu, Anzhe Xie, Changyue Wang, Yixiao Ma, Haitao Li, Zhijing Wu, Yiqun Liu, and Min Zhang. 2025. Pre-training for legal case retrieval based on inter-case distinctions.ACM Transactions on Information Systems43, 5 (2025), 1–27

    2025

  24. [24]

    Weihang Su, Xuanyi Chen, Yueyue Wu, Qingyao Ai, and Yiqun Liu. 2026. Enhanc- ing Judgment Document Generation via Agentic Legal Information Collection and Rubric-Guided Optimization.arXiv preprint arXiv:2605.02011(2026)

    Pith/arXiv arXiv 2026

  25. [25]

    Weihang Su, Yiran Hu, Anzhe Xie, Qingyao Ai, Quezi Bing, Ning Zheng, Yun Liu, Weixing Shen, and Yiqun Liu. 2024. STARD: A Chinese Statute Retrieval Dataset Derived from Real-life Queries by Non-professionals. InFindings of the Association for Computational Linguistics: EMNLP 2024. 10658–10671

    2024

  26. [26]

    Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024. Dragin: Dynamic retrieval augmented generation based on the real-time information needs of large language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 12991–13013

    2024

  27. [27]

    Weihang Su, Yichen Tang, Qingyao Ai, Junxi Yan, Changyue Wang, Hongning Wang, Ziyi Ye, Yujia Zhou, and Yiqun Liu. 2025. Parametric retrieval augmented generation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1240–1250

    2025

  28. [28]

    Weihang Su, Baoqing Yue, Qingyao Ai, Yiran Hu, Jiaqi Li, Changyue Wang, Kaiyuan Zhang, Yueyue Wu, and Yiqun Liu. 2025. Judge: Benchmarking judg- ment document generation for chinese legal system. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 3573–3583

    2025

  29. [29]

    Yi Tay, Vinh Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, et al. 2022. Transformer memory as a differentiable search index.Advances in Neural Information Processing Systems 35 (2022), 21831–21843

    2022

  30. [30]

    Qwen Team et al. 2024. Qwen2 technical report.arXiv preprint arXiv:2407.10671 2, 3 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

    Marc Van Opijnen and Cristiana Santos. 2017. On the concept of relevance in legal information retrieval.Artificial Intelligence and Law25, 1 (2017), 65–87

    2017

  32. [32]

    Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, et al . 2022. A neural corpus indexer for document retrieval.Advances in Neural Information Processing Systems35 (2022), 25600–25614

    2022

  33. [33]

    Zihan Wang, Yujia Zhou, Yiteng Tu, and Zhicheng Dou. 2023. NOVO: learnable and interpretable document identifiers for model-based IR. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2656–2665

    2023

  34. [34]

    Gineke Wiggers, Suzan Verberne, Wouter van Loon, and Gerrit-Jan Zwenne

  35. [35]

    citations as flavors of impact relevance.Journal of the Association for Information Science and Technology74, 8 (2023), 1010–1025

    Bibliometric-enhanced legal information retrieval: Combining usage and Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Yiteng Tu et al. citations as flavors of impact relevance.Journal of the Association for Information Science and Technology74, 8 (2023), 1010–1025

    2018

  36. [36]

    Gineke Wiggers, Suzan Verberne, Gerrit-Jan Zwenne, and Wouter Van Loon

  37. [37]

    Exploration of domain relevance by legal professionals in information retrieval systems.Legal Information Management22, 1 (2022), 49–67

    2022

  38. [38]

    Chaojun Xiao, Xueyu Hu, Zhiyuan Liu, Cunchao Tu, and Maosong Sun. 2021. Lawformer: A pre-trained language model for chinese legal long documents.AI Open2 (2021), 79–84

    2021

  39. [39]

    Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-pack: Packed resources for general chinese embeddings. InProceedings of the 47th international ACM SIGIR conference on research and development in information retrieval. 641–649

    2024

  40. [40]

    Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, et al . 2021. mT5: A massively multilingual pre-trained text-to-text transformer. InProceedings of the 2021 conference of the North American chapter of the association for computational linguistics: Human language technologies. 483–498

    2021

  41. [41]

    Mingruo Yuan, Ben Kao, and Tien-Hsuan Wu. 2025. QBR: A Question-Bank- Based Approach to Fine-Grained Legal Knowledge Retrieval for the General Public.arXiv preprint arXiv:2505.04883(2025)

    arXiv 2025

  42. [42]

    Yujia Zhou, Jing Yao, Zhicheng Dou, Ledell Wu, Peitian Zhang, and Ji-Rong Wen

  43. [43]

    Ultron: An ultimate retriever on corpus with a model-based indexer.arXiv preprint arXiv:2208.09257(2022)

    Pith/arXiv arXiv 2022

  44. [44]

    Shengyao Zhuang, Houxing Ren, Linjun Shou, Jian Pei, Ming Gong, Guido Zuc- con, and Daxin Jiang. 2022. Bridging the gap between indexing and retrieval for differentiable search index with query generation.arXiv preprint arXiv:2206.10128 (2022)

    Pith/arXiv arXiv 2022