Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Regulation-driven fine-grained classification needs constraint-aware hierarchical search, not flat semantic matching, and a tree-local next-hop method tops LLM and RAG baselines on four regulatory benchmarks.

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 10:37 UTC pith:TCJTECT6

load-bearing objection Solid systems paper: clean task framing, four expert-checked benchmarks, and real gains from local hierarchical search over strong RAG baselines; extraction completeness is a soft spot but not load-bearing for the claim. the 3 major comments →

arxiv 2607.10588 v1 pith:TCJTECT6 submitted 2026-07-12 cs.AI cs.CL

Constraint-Aware Hierarchical Search for Regulation-Driven Fine-Grained Classification

Siyu Wang , Wei Tan , Lulu Chen This is my paper
classification cs.AI cs.CL
keywords regulation-driven classificationhierarchical rule-based reasoningconstraint-aware searchfine-grained hierarchical classificationretrieval-augmented generationcustoms tariff classificationexport control categorization
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.

In customs tariffs, export controls, and equipment coding, the correct label is fixed by explicit rules—thresholds, exclusions, definitions, and local exceptions—not by how similar an input looks to a category description. Two nearly identical products can require different codes, and a retrieved passage can be relevant yet legally out of scope. The authors formulate this as regulation-driven fine-grained hierarchical classification: assign an external instance to a fine-grained class only along a valid path in a regulatory hierarchy, backed by auditable evidence. They build four expert-validated benchmarks and a framework that turns regulations into a searchable tree, retrieves only valid child candidates at each step, and chooses the next hop using structured rule fields plus evidence snippets. The method achieves the best mean accuracy on all four datasets and returns interpretable paths, with the largest gains where neighboring categories are separated by small rule conditions.

Core claim

The paper claims that regulation-driven classification cannot be solved reliably by flat classifiers, conventional hierarchical text models, or unconstrained retrieval-augmented LLMs. It must be cast as constrained search over a regulatory hierarchy, and a system that retrieves only valid local candidates and exposes notes, exclusions, definitions, and evidence at each next-hop decision outperforms direct LLM and several RAG baselines on final-label accuracy while producing hierarchy-consistent paths and supporting evidence.

What carries the argument

Constraint-aware hierarchical search: offline conversion of regulatory documents into a searchable tree with node and evidence indexes; online local retrieval of only valid child candidates under the current node; candidate packages that include code, title, text, notes, exclusions, definitions, and evidence snippets; LLM next-hop selection under those constraints, with optional fast anchor initialization and path-level evidence aggregation for final verification.

Load-bearing premise

The method assumes an offline LLM can extract a complete and correct regulatory tree—nested children, notes, exclusions, and definitions—from the source documents; if that tree is incomplete or wrong, local search runs on a broken decision space.

What would settle it

On the same HS-Hard and Dual-Use test sets with the same embedding and LLM backend, if a carefully tuned flat evidence RAG with candidate reranking matched or exceeded the reported final-label accuracies while also producing valid hierarchy paths, the claim that hierarchy-scoped local search is required for those gains would fail.

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

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 formulates regulation-driven fine-grained hierarchical classification: assign an external instance to a fine-grained class via a valid path in a regulatory hierarchy, under rule-defined boundaries (thresholds, exclusions, definitions), with auditable evidence. It constructs four expert-validated benchmarks (HS-Simple, HS-Hard, Dual-Use List, Educational Equipment) and proposes a constraint-aware hierarchical search framework that offline builds a searchable regulatory tree with node/evidence indexes, then online performs local child retrieval, candidate-package construction (code/title/text/notes/exclusions/definitions/evidence), LLM next-hop decisions, path-level evidence aggregation, and optional fast-anchor initialization. Experiments against six LLM/RAG baselines (direct LLM, Flat Evidence RAG + rerank, LightRAG, GraphRAG, LinearRAG, MA-RAG) report best mean final-label accuracy on all four datasets over three runs (Table 3), with level-wise HS analysis (Table 4), ablations on anchor, topM, and candidate fields (Figures 3–5), and two HS-Hard case studies isolating subheading rules.

Significance. If the results hold, the work is a solid systems contribution for high-stakes regulatory classification (customs, export control, standards coding). Strengths include: (i) a clear task formulation that separates hierarchical validity and rule consistency from pure semantic matching; (ii) four expert-in-the-loop benchmarks with explicit hierarchy paths; (iii) a reproducible framework (code and data linked) that enforces parent–child constraints and attaches structured regulatory fields; (iv) multi-run means and sample standard deviations, level-wise accuracy, three ablations, and qualitative cases that isolate decisive local rules. Gains are largest where neighboring fine-grained categories and boundary conditions matter (HS-Hard +10.36 pp, Educational Equipment +10.56 pp over strongest baselines), which is the practically important regime. The contribution is empirical and engineering-oriented rather than a new theoretical result, but it is well-scoped and useful for the field.

major comments (3)
  1. Section 3.2 and Appendix A: the offline LLM extraction of nested children, notes, exclusions, and definitions is load-bearing for the searchable decision space, yet the paper reports no quantitative extraction quality (node/edge precision-recall, missing-note rate, or expert audit of the built trees). Residual risk is real for HS-Hard (final Acc 58.29%). Please add a short tree-quality audit (sample of nodes/edges against source documents) or an ablation that freezes the same tree for all hierarchy-aware variants so residual extraction error cannot selectively favor the proposed method.
  2. Section 4.2 / Table 3: baselines are exclusively LLM/RAG systems; classical hierarchical text classification (local classifiers, hierarchy-aware neural models cited in §2.1) and supervised fine-grained classifiers are omitted. The authors argue labeled data are scarce and regulations update, which is reasonable, but at least one hierarchy-aware supervised or multi-label baseline on the same tree (or a frozen-label flat classifier over terminal codes) would strengthen the claim that constraint-aware search, not merely access to the hierarchy, drives the gains—especially on Educational Equipment where accuracy is already high.
  3. Section 4.3.1: the final generator G_θ is designed to flag low-confidence / insufficient-evidence cases, yet all benchmark instances are treated as answerable and abstention is left to future work. For a regulation-driven claim that emphasizes auditability and not over-committing, a minimal report of confidence calibration or forced-abstention behavior on a held-out ambiguous subset (or synthetic missing-attribute variants of Table 2) would better support the safety-oriented part of the contribution.
minor comments (5)
  1. Table 1: HS-Simple and HS-Hard both list 370 instances; Dual-Use 270 and Educational Equipment 300. Please clarify train/dev/test splits (or confirm pure zero-shot evaluation) and how difficulty scores (2–5) were assigned.
  2. Figure 5 and §4.6.3: candidate-field ablation numbers are described in text but absolute accuracies for Title-only / Title+Evidence / Title+Text / Full are not tabulated; a small table would aid reproducibility.
  3. Equation (16) and Algorithm 1: the stop signal σ_t, confidence ρ_t, and rationale η_t are produced by F_θ, but the prompt schema and parsing rules for these fields are only sketched in Appendix A; a short example I/O would help.
  4. Related work §2.3–2.4: several 2024–2025 RAG systems are cited; a one-sentence contrast with hierarchical document refinement / hierarchical RAG (e.g., Jin et al. 2025 already in the bibliography) would tighten positioning.
  5. Minor presentation: occasional missing spaces after periods in the abstract and introduction (e.g., “whilearetrievedpassage”); fix for camera-ready.

Circularity Check

0 steps flagged

No significant circularity; empirical accuracy claims are measured against independent expert-held-out labels and do not reduce by construction to method inputs.

full rationale

The paper formulates regulation-driven fine-grained hierarchical classification, constructs four expert-validated benchmarks, and proposes a constraint-aware hierarchical search that builds a regulatory tree offline then performs local next-hop decisions. Its central claims (best mean final-label accuracy on all four datasets, largest gains on boundary-sensitive cases) are purely empirical comparisons against gold codes and paths that are fixed by source regulations and expert-in-the-loop annotation, independent of the method's retrieval scores, candidate packages, or LLM decisions. No equation equates a predicted quantity to a fitted input by construction; hyperparameters (topM=8, max depth=8) are chosen on development data in the ordinary way and do not force test-set identities. There are no self-citations of uniqueness theorems, no ansatz imported from prior author work that defines the target result, and no renaming of a known pattern presented as a derived unification. Tree extraction completeness is an assumption about implementation fidelity, not a circular reduction of the reported accuracies. The evaluation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

As an empirical systems paper the central claim rests on standard ML evaluation assumptions plus a small set of design choices (candidate size, LLM extraction fidelity, expert annotation process). No new physical entities or free parameters that define the target accuracy are introduced; the free parameters are ordinary hyper-parameters of the search procedure.

free parameters (3)
  • topM (candidate set size) = 8
    Controls both global anchor retrieval and local child retrieval; set to 8 after development-set search and shown to affect HS-Hard accuracy (Figure 4).
  • maximum search depth Dmax = 8
    Hard stop for the hierarchical traversal; fixed at 8 in the implementation appendix.
  • evidence aggregation limits (chunks per node / per pass / total) = 2/6/10
    Up to 2 chunks per node, 6 per pass, 10 total before final generator; chosen by authors for context length.
axioms (4)
  • domain assumption An LLM can extract nested children, notes, exclusions and definitions from regulatory text blocks with sufficient fidelity to support later local search.
    Invoked in Section 3.2 and Appendix A; the entire offline tree is produced by this extraction step.
  • domain assumption Expert-in-the-loop validation yields gold paths and labels that correctly reflect the official regulatory hierarchy and decisive conditions.
    Stated in Section 4.1; all accuracy numbers are measured against these labels.
  • domain assumption Cosine similarity of text-embedding-v4 vectors is a useful ranking signal for retrieving plausible child nodes and evidence snippets inside a regulatory scope.
    Used throughout Algorithm 1 and the experimental setup.
  • standard math Standard hierarchical classification and RAG evaluation metrics (exact final-label match, level-wise accuracy) are appropriate for the new task.
    Equations (22)–(23) and Table 3.
invented entities (2)
  • Regulation-driven Fine-grained Hierarchical Classification (task paradigm) no independent evidence
    purpose: Unifies customs, export-control and standards coding as a single constrained search problem requiring hierarchy validity, rule consistency and boundary reasoning.
    Introduced in the abstract and Section 1 as a named contribution; no independent prior benchmark of this exact form is cited.
  • Constraint-aware Hierarchical Search Framework (tree + local packages + next-hop LLM) no independent evidence
    purpose: Operationalizes the task by converting regulations into a searchable tree and performing only valid local next-hop decisions enriched with structured regulatory fields.
    Core technical proposal of Sections 3.2–3.3; evaluated empirically but not independently verified outside the paper’s own experiments.

pith-pipeline@v1.1.0-grok45 · 26100 in / 2873 out tokens · 39642 ms · 2026-07-14T10:37:09.589763+00:00 · methodology

0 comments
read the original abstract

Tasks such as customs tariff classification, export control categorization, and standards-based equipment coding require assigning an input instance to a fine-grained class under an explicit regulatory hierarchy. Unlike standard text classification, the correct label in these tasks is not determined by semantic similarity alone, but by rule-defined boundaries, threshold conditions, exclusion clauses, definitions, and local exceptions. As a result, two highly similar inputs may require different labels, while a retrieved passage that appears relevant may still be inapplicable under the governing rules. Existing flat classifiers, hierarchical text classification methods, and retrieval-augmented LLM systems are not designed to jointly enforce hierarchical validity, rule consistency, and fine-grained boundary reasoning. In this paper, we formulate this setting as regulation-driven fine-grained hierarchical classification, where an external instance must be assigned to a fine-grained class through a valid path in a regulatory hierarchy and supported by auditable evidence. We construct four benchmark datasets from representative regulation-intensive scenarios and validate the annotations through an expert-in-the-loop process. We further propose a constraint-aware hierarchical search framework that converts regulatory documents into a searchable tree, retrieves only valid local candidate nodes, and uses structured regulatory fields with evidence snippets to guide each next-hop decision. Experiments show that our method achieves the best mean accuracy on all four datasets and provides interpretable decision paths, with the largest gains on cases involving fine-grained neighboring categories and rule-based boundary conditions.

Figures

Figures reproduced from arXiv: 2607.10588 by Lulu Chen, Siyu Wang, Wei Tan.

Figure 1
Figure 1. Figure 1: A motivating example from customs tariff classification. A semantic match may retrieve a plausible but invalid code, while a regulation-driven decision must follow the hierarchy and check the decisive conditions. with only spark-ignition engines. It may also recognize the hybrid signal but predict HS code 8703.40, which covers hybrid vehicles that are not externally chargeable. Both predictions are plausib… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed framework. Offline, regulatory documents are converted into a searchable tree with node and evidence indexes. Online, the model searches the tree step by step, compares only valid local candidates, and aggregates evidence along the selected path. 3. Constraint-Aware Hierarchical Search Framework Our method has two stages, as shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation study on HS-Hard comparing Fast Anchor initialization with a baseline without it. Fast Anchor improves final-label accuracy from 50.54% to 58.29% and reduces the average search path length from 2.89 to 1.55, indicating both higher classification performance and more efficient search [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity of candidate set size 𝑡𝑜𝑝𝑀 on HS bench￾marks. HS-Simple shows limited sensitivity to 𝑡𝑜𝑝𝑀 (87.90%, 86.83%, and 90.44% for 𝑡𝑜𝑝𝑀 = 3, 5, and 8), while HS-Hard benefits from larger 𝑡𝑜𝑝𝑀 (54.59% → 58.29%), suggesting that harder instances require higher candidate coverage to keep the correct next hop among candidates [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Candidate field ablation for next-hop branch selection (𝐹𝜃 ) with 𝑡𝑜𝑝𝑀 = 8. We progressively enrich the candidate package Γ(𝑢) from Title-only to Full Candidate and report final-label accuracy on HS-Hard and HS-Simple. The first three variants omit explicit regulatory fields such as notes, exclusions, and definitions, while the Full Candidate variant includes them. fine-grained neighboring codes and explic… view at source ↗

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

40 extracted references · 8 linked inside Pith

  1. [1]

    and Freitas, Alex A

    Silla, Carlos N. and Freitas, Alex A. , title =. Data Mining and Knowledge Discovery , year =

  2. [2]

    and Heidarysafa, Mojtaba and Jafari Meimandi, Kiana and Gerber, Matthew S

    Kowsari, Kamran and Brown, Donald E. and Heidarysafa, Mojtaba and Jafari Meimandi, Kiana and Gerber, Matthew S. and Barnes, Laura E. , title =. Proceedings of the 16th IEEE International Conference on Machine Learning and Applications , year =

  3. [3]

    , title =

    Wehrmann, Jonatas and Cerri, Ricardo and Barros, Rodrigo C. , title =. Proceedings of the 35th International Conference on Machine Learning , year =

  4. [4]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

    Zhou, Jie and Ma, Chunping and Long, Dingkun and Xu, Guangwei and Ding, Ning and Zhang, Haoyu and Xie, Pengjun and Liu, Gongshen , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

  5. [5]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , year =

    Wang, Zihan and Wang, Peiyi and Huang, Lianzhe and Sun, Xin and Wang, Houfeng , title =. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , year =

  6. [6]

    , title =

    Ling, Xiao and Weld, Daniel S. , title =. Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence , year =

  7. [7]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics , year =

    Choi, Eunsol and Levy, Omer and Choi, Yejin and Zettlemoyer, Luke , title =. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics , year =

  8. [8]

    and Ji, Heng and Han, Jiawei , title =

    Ren, Xiang and He, Wenqi and Qu, Meng and Voss, Clare R. and Ji, Heng and Han, Jiawei , title =. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , year =

    2016

  9. [9]

    Proceedings of the 5th Workshop on Structured Prediction for NLP , year =

    Dayanik, Erenay and Blessing, Andre and Blokker, Nico and Haunss, Sebastian and Kuhn, Jonas and Lapesa, Gabriella and Pado, Sebastian , title =. Proceedings of the 5th Workshop on Structured Prediction for NLP , year =

  10. [10]

    Findings of the Association for Computational Linguistics: EMNLP 2020 , year =

    Chalkidis, Ilias and Fergadiotis, Manos and Malakasiotis, Prodromos and Aletras, Nikolaos and Androutsopoulos, Ion , title =. Findings of the Association for Computational Linguistics: EMNLP 2020 , year =

    2020

  11. [11]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , year =

    Chalkidis, Ilias and Jana, Abhik and Hartung, Dirk and Bommarito, Michael and Androutsopoulos, Ion and Katz, Daniel Martin and Aletras, Nikolaos , title =. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , year =

  12. [12]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

    Zhong, Haoxi and Xiao, Chaojun and Tu, Cunchao and Zhang, Tianyang and Liu, Zhiyuan and Sun, Maosong , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

  13. [13]

    arXiv preprint arXiv:2103.06268 , year =

    Hendrycks, Dan and Burns, Collin and Chen, Anya and Ball, Spencer , title =. arXiv preprint arXiv:2103.06268 , year =

    Pith/arXiv arXiv

  14. [14]

    Guha, Neel and Nyarko, Julian and Ho, Daniel E. and R. arXiv preprint arXiv:2308.11462 , year =

    Pith/arXiv arXiv

  15. [15]

    PLOS ONE , year =

    Moreira Valle, Let. PLOS ONE , year =

  16. [16]

    A Data Decomposition-Based Hierarchical Classification Method for Multi-Label Classification of Contractual Obligations for the Purpose of Their Governance , journal =

    Singh, Amrita and Anish, Preethu Rose and Verma, Aparna and Venkatesan, Sivanthy and. A Data Decomposition-Based Hierarchical Classification Method for Multi-Label Classification of Contractual Obligations for the Purpose of Their Governance , journal =. 2024 , doi =

    2024

  17. [17]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and Kuttler, Heinrich and Lewis, Mike and Yih, Wen-tau and Rockt. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , year =

  18. [18]

    Proceedings of the 37th International Conference on Machine Learning , year =

    Guu, Kelvin and Lee, Kenton and Tung, Zora and Pasupat, Panupong and Chang, Ming-Wei , title =. Proceedings of the 37th International Conference on Machine Learning , year =

  19. [19]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , year =

    Karpukhin, Vladimir and Oguz, Barlas and Min, Sewon and Lewis, Patrick and Wu, Ledell and Edunov, Sergey and Chen, Danqi and Yih, Wen-tau , title =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , year =

    2020

  20. [20]

    Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics , year =

    Izacard, Gautier and Grave, Edouard , title =. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics , year =

  21. [21]

    Proceedings of the 11th International Conference on Learning Representations , year =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , title =. Proceedings of the 11th International Conference on Learning Representations , year =

  22. [22]

    arXiv preprint arXiv:2312.10997 , year =

    Gao, Yunfan and Xiong, Yun and Gao, Xinyu and Jia, Kangxiang and Pan, Jinliu and Bi, Yuxi and Dai, Yi and Sun, Jiawei and Wang, Meng and Wang, Haofen , title =. arXiv preprint arXiv:2312.10997 , year =

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2404.10981 , year =

    Huang, Yizheng and Huang, Jimmy , title =. arXiv preprint arXiv:2404.10981 , year =

    Pith/arXiv arXiv

  24. [24]

    arXiv preprint arXiv:2404.16130 , year =

    Edge, Darren and Trinh, Ha and Cheng, Newman and Bradley, Joshua and Chao, Alex and Mody, Apurva and Truitt, Steven and Metropolitansky, Dasha and Osazuwa Ness, Robert and Larson, Jonathan , title =. arXiv preprint arXiv:2404.16130 , year =

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2410.05779 , year =

    Guo, Zirui and Xia, Lianghao and Yu, Yanhua and Ao, Tu and Huang, Chao , title =. arXiv preprint arXiv:2410.05779 , year =

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2510.10114 , year =

    Zhuang, Luyao and Chen, Shengyuan and Xiao, Yilin and Zhou, Huachi and Zhang, Yujing and Chen, Hao and Zhang, Qinggang and Huang, Xiao , title =. arXiv preprint arXiv:2510.10114 , year =

    arXiv

  27. [27]

    arXiv preprint arXiv:2505.20096 , year =

    Nguyen, Thang and Chin, Peter and Tai, Yu-Wing , title =. arXiv preprint arXiv:2505.20096 , year =

    arXiv

  28. [28]

    2: Pushing the frontier of open large language models , author=

    Deepseek-v3. 2: Pushing the frontier of open large language models , author=. arXiv preprint arXiv:2512.02556 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2602.03442 , year=

    A-RAG: Scaling Agentic Retrieval-Augmented Generation via Hierarchical Retrieval Interfaces , author=. arXiv preprint arXiv:2602.03442 , year=

    arXiv

  30. [30]

    arXiv preprint arXiv:2407.08959 , year=

    Domain-hierarchy adaptation via chain of iterative reasoning for few-shot hierarchical text classification , author=. arXiv preprint arXiv:2407.08959 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2510.01474 , year=

    AIReg-Bench: Benchmarking language models that assess AI regulation compliance , author=. arXiv preprint arXiv:2510.01474 , year=

    arXiv

  32. [32]

    Knowledge graph-guided retrieval augmented generation , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  33. [33]

    arXiv preprint arXiv:2503.10150 , year=

    Retrieval-augmented generation with hierarchical knowledge , author=. arXiv preprint arXiv:2503.10150 , year=

    arXiv

  34. [34]

    International conference on learning representations , volume=

    Self-rag: Learning to retrieve, generate, and critique through self-reflection , author=. International conference on learning representations , volume=

  35. [35]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Atm: Adversarial tuning multi-agent system makes a robust retrieval-augmented generator , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  36. [36]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    HYDRA: A Multi-Head Encoder-only Architecture for Hierarchical Text Classification , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  37. [37]

    Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    EnCore: fine-grained entity typing by pre-training entity encoders on coreference chains , author=. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  38. [38]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Ontotype: Ontology-guided and pre-trained language model assisted fine-grained entity typing , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  39. [39]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    IL-TUR: Benchmark for Indian legal text understanding and reasoning , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  40. [40]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Hierarchical document refinement for long-context retrieval-augmented generation , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=