Pith. sign in

REVIEW 3 major objections 5 minor 20 references

Faithful or Findable? Evaluating LLM-Generated Metadata for RDF Dataset Search

T0 review · 3 major / 5 minor · reviewed 2026-07-08 · grok-4.5

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

Pith's one-line read Unconstrained LLM rewriting of RDF dataset metadata improves search most but expands meaning the data do not support; profile-grounded rewriting balances effectiveness and faithfulness.

desk verdict Useful dual-objective framing for LLM RDF metadata: unconstrained rewriting wins retrieval and loses faithfulness; the causal “unsupported expansion” story is the soft spot until metrics are checked. read the letter →

arxiv 2607.05970 v1 pith:6ZLHS73Q submitted 2026-07-07 cs.IR cs.AI

Riccardo Terrenzi , Serkan Ayvaz This is my paper · ORCID
classification cs.IRcs.AI
keywords LLM-generatedmetadataRDFdatasetsdatasetsearchretrievaleffectivenessfaithfulnesssyntheticrewritinginformation
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

Dataset search depends on metadata, so LLM-generated descriptions of RDF datasets can change what users find. This paper tests six generation settings, from free rewriting of existing metadata through profile-grounded and agentic graph-based methods, and scores them jointly on retrieval effectiveness and faithfulness. Unconstrained rewriting produces the largest gains over original metadata, yet it is also the least faithful: the search lift can come from unsupported semantic expansion rather than accurate description. Grounded settings recover much of the faithfulness while still improving retrieval, and profile-grounded rewriting gives the most balanced trade-off. The result frames synthetic metadata as a system-level IR problem in which effectiveness, provenance, and trust must be measured together.

What carries the argument

Joint evaluation of retrieval effectiveness and faithfulness across six metadata-generation settings for RDF datasets—from unconstrained rewriting to profile-grounded and agentic graph-based generation—used to isolate how much of any retrieval gain is purchased by unsupported semantic expansion versus grounded description.

What would settle it

Re-annotate a sample of passages the metric labels unfaithful against the original RDF graphs and profiles; if most expansions are actually entailed, or if removing the expanded terms erases the retrieval advantage of unconstrained rewriting, the central claim fails.

Watch

Extended reading notes

Core claim

Unconstrained LLM rewriting of RDF dataset metadata delivers the strongest retrieval gains relative to the original metadata, but it is also the least faithful, showing that search improvements can be driven by unsupported semantic expansion. More grounded settings substantially raise faithfulness, and profile-grounded rewriting supplies the most balanced trade-off between retrieval effectiveness and grounding.

Load-bearing premise

The paper’s faithfulness metric correctly flags unsupported semantic expansion rather than legitimate clarification of latent but true RDF properties, and that this expansion is what causes unconstrained rewriting’s retrieval gains.

Editorial extensions

If this is right

  • Search systems that accept unconstrained LLM-rewritten metadata without faithfulness checks can rank datasets on invented properties.
  • Profile-grounded rewriting is a practical default that improves findability while limiting ungrounded expansion.
  • Dataset portals and RDF repositories need joint metrics for effectiveness, provenance, and trust when deploying synthetic metadata.
  • Agentic graph-based generation is a route to higher faithfulness when full structural access is available.
  • Evaluation of synthetic content in IR must treat faithfulness as a first-class criterion alongside traditional relevance metrics.

Reading between the lines

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

  • Part of the retrieval gain from unconstrained rewriting may be an artifact of query–document vocabulary matching rather than true semantic enrichment of the dataset.
  • The same effectiveness-versus-faithfulness tension is likely to appear in other metadata-heavy domains such as scientific data repositories and enterprise data catalogs.
  • A direct user study with provenance indicators would show whether people prefer the more faithful but slightly less effective metadata or the unconstrained version.
  • If the faithfulness metric under-penalizes latent but true properties, the paper may overstate the unfaithfulness of unconstrained rewriting; re-checking expansions against the graphs would settle it.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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. This paper studies six LLM-based metadata-generation settings for RDF datasets—ranging from unconstrained rewriting through profile-grounded rewriting to agentic graph-based generation—and evaluates them jointly for retrieval effectiveness and faithfulness to the underlying data. The central empirical claim is that unconstrained rewriting yields the strongest retrieval gains over original metadata while being the least faithful, which the authors interpret as evidence that search improvements can be driven by unsupported semantic expansion; more grounded settings improve faithfulness, and profile-grounded rewriting is presented as the best effectiveness–faithfulness trade-off. The work frames synthetic metadata as a system-level IR problem in which effectiveness, provenance, and trust must be assessed together.

Significance. Jointly measuring retrieval effectiveness and faithfulness for LLM-generated dataset metadata is a timely and practically relevant contribution for IR and semantic-web dataset search. If the comparative rankings and the causal interpretation hold, the paper supplies actionable guidance (favoring profile-grounded methods for balanced performance) and a useful framing that synthetic enrichment of search corpora cannot be judged by effectiveness alone. The multi-condition design and dual-metric evaluation are genuine strengths. The manuscript does not appear to ship machine-checked proofs or parameter-free derivations; its value is empirical and conceptual. That value is contingent on the soundness of the faithfulness operationalization and on isolating unsupported expansion from confounds—points that currently limit how strongly the central claim can be endorsed.

major comments (3)
  1. [Abstract; faithfulness metric definition and results discussion] The load-bearing interpretive claim—that unconstrained rewriting’s retrieval gains are driven by unsupported semantic expansion—requires that the faithfulness metric correctly flags content unsupported by the RDF graph, as opposed to legitimate clarification of latent but true properties, SPARQL/RDFS/OWL-entailed facts, or profile-implied structure. Surface-overlap or direct-triple-presence metrics systematically under-credit valid graph-derived summaries. The manuscript needs either (i) validation of faithfulness labels against an entailment-aware oracle or calibrated human judgments, or (ii) an explicit statement of the metric’s limitations for RDF and a corresponding softening of the causal gloss. Without this, the descriptive ranking of settings may stand while the “showing that” clause does not.
  2. [Experimental design / results (retrieval vs. faithfulness joint analysis)] Even if faithfulness labels are correct, retrieval lift for unconstrained rewriting can be produced by confounds (length, lexical diversity, query-term coverage from better writing) rather than specifically by unsupported semantic content. The paper should report controls or ablations—e.g., length-matched or vocabulary-controlled baselines, or experiments that strip only non-entailed spans and re-measure retrieval—to isolate the contribution of unsupported expansion. Absent such isolation, the causal attribution remains the least secure step in the argument, even if the effectiveness–faithfulness ranking of the six settings is robust.
  3. [Results / trade-off claim for profile-grounded rewriting] The claim that profile-grounded rewriting provides the “most balanced trade-off” needs an explicit decision rule or multi-objective summary (e.g., Pareto front, weighted score with stated weights, or constrained optimization). Without a transparent aggregation of the two metrics across the six settings, “most balanced” is a qualitative gloss rather than a reproducible finding. A table or figure that makes the trade-off criterion inspectable would make this central recommendation falsifiable and comparable across follow-up work.
minor comments (5)
  1. [Abstract] The abstract states clear comparative outcomes but would benefit from brief quantitative anchors (relative retrieval gains and faithfulness score ranges) so readers can gauge effect sizes before the full results section.
  2. [Methods / experimental settings] Name and briefly define all six generation settings in a single early table or enumerated list so that later references to “unconstrained,” “profile-grounded,” and “agentic” settings are unambiguous.
  3. [Experimental setup] Report dataset sizes, query-set construction, number of RDF graphs, and any statistical tests or error bars for retrieval and faithfulness differences; these are standard for IR comparative claims and are not visible from the abstract alone.
  4. [Reproducibility / artifacts] If prompts, generation configurations, query sets, and evaluation scripts are not already released, doing so would substantially strengthen reproducibility and allow others to stress-test the faithfulness operationalization.
  5. [Preliminaries / metrics] Clarify notation and terminology for “faithfulness,” “grounding,” and “unsupported semantic expansion” at first use so that the dual-metric design is readable without reconstructing definitions from results prose.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive reading. The three major comments correctly identify where our interpretive claims outrun the current operationalization of faithfulness and the isolation of confounds, and where the trade-off recommendation needs a transparent decision rule. We accept these points and will revise accordingly: (1) we will state the faithfulness metric’s limitations for RDF entailment and soften the causal gloss that unconstrained gains are driven by unsupported expansion; (2) we will add length- and vocabulary-oriented controls and clarify residual confounds; (3) we will make the effectiveness–faithfulness trade-off criterion explicit (Pareto-style summary and a stated aggregation). These changes preserve the descriptive ranking of the six settings while making the central claims falsifiable and appropriately scoped.

read point-by-point responses

  1. Referee: The load-bearing interpretive claim—that unconstrained rewriting’s retrieval gains are driven by unsupported semantic expansion—requires that the faithfulness metric correctly flags content unsupported by the RDF graph, as opposed to legitimate clarification of latent but true properties, SPARQL/RDFS/OWL-entailed facts, or profile-implied structure. Surface-overlap or direct-triple-presence metrics systematically under-credit valid graph-derived summaries. The manuscript needs either (i) validation of faithfulness labels against an entailment-aware oracle or calibrated human judgments, or (ii) an explicit statement of the metric’s limitations for RDF and a corresponding softening of the causal gloss. Without this, the descriptive ranking of settings may stand while the “showing that” clause does not.

    Authors: We agree. Our faithfulness operationalization is grounded in surface/profile overlap and direct support relative to the RDF graph and profiles; it does not implement an entailment-aware oracle (RDFS/OWL/SPARQL-level) and therefore can under-credit legitimate graph-derived or profile-implied content. We will not claim that the metric fully separates unsupported invention from valid entailment. In revision we will: (a) add an explicit Limitations subsection stating that the metric is a conservative, non-entailment-aware proxy and may penalize valid clarification; (b) soften the Abstract and Results language from “showing that search improvements can be driven by unsupported semantic expansion” to a more careful formulation (e.g., that unconstrained rewriting yields the largest gains while scoring lowest on our support metric, consistent with—but not proving—unsupported expansion); and (c) where space allows, report a small calibrated human spot-check on a sample of flagged spans to illustrate agreement/disagreement patterns. Full entailment-oracle validation across all datasets is beyond the present revision scope; option (ii) is the path we take. The descriptive ranking of settings remains; the causal “showing that” clause will be qualified. revision: yes

  2. Referee: Even if faithfulness labels are correct, retrieval lift for unconstrained rewriting can be produced by confounds (length, lexical diversity, query-term coverage from better writing) rather than specifically by unsupported semantic content. The paper should report controls or ablations—e.g., length-matched or vocabulary-controlled baselines, or experiments that strip only non-entailed spans and re-measure retrieval—to isolate the contribution of unsupported expansion. Absent such isolation, the causal attribution remains the least secure step in the argument, even if the effectiveness–faithfulness ranking of the six settings is robust.

    Authors: The referee is right that length, lexical diversity, and query-term coverage can confound attribution of retrieval lift to unsupported semantic content. We will strengthen isolation as follows. First, we will report length statistics and length-matched or length-normalized analyses (e.g., truncating or sampling unconstrained outputs to the length distribution of grounded settings, and/or correlating effectiveness with length within setting). Second, we will report simple vocabulary/diversity controls (type–token and query-term coverage) and discuss residual lift after accounting for them. Third, we will clarify in the text that we do not claim a fully causal isolation of “unsupported expansion” as the sole driver; the joint effectiveness–faithfulness ranking is the primary empirical result, and any causal gloss will be presented as a hypothesis consistent with the pattern rather than a demonstrated mechanism. A full “strip only non-entailed spans and re-retrieve” ablation would require a reliable entailment partition of every generated span, which we do not have (see Comment 1); we therefore treat that ablation as out of scope and state the residual confound explicitly. These additions make the security of the causal step transparent without overstating what the design can isolate. revision: partial

  3. Referee: The claim that profile-grounded rewriting provides the “most balanced trade-off” needs an explicit decision rule or multi-objective summary (e.g., Pareto front, weighted score with stated weights, or constrained optimization). Without a transparent aggregation of the two metrics across the six settings, “most balanced” is a qualitative gloss rather than a reproducible finding. A table or figure that makes the trade-off criterion inspectable would make this central recommendation falsifiable and comparable across follow-up work.

    Authors: We accept this fully. “Most balanced” was a qualitative reading of the joint plot and should be replaced by an inspectable rule. In revision we will: (1) present the six settings in the effectiveness–faithfulness plane and mark the Pareto front; (2) report a simple, stated multi-objective summary (e.g., min–max normalized scores and a small set of fixed weights, plus a constrained view such as “best effectiveness among settings above a faithfulness threshold”); and (3) revise Abstract/Results wording so that the recommendation for profile-grounded rewriting is tied to that explicit criterion rather than an informal gloss. This makes the trade-off claim reproducible and comparable for follow-up work. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical comparison of generation settings against external retrieval and faithfulness metrics; no derivation reduces a prediction to a fitted input or self-definition.

full rationale

The paper is an empirical IR evaluation of six LLM metadata-generation settings for RDF datasets. It measures retrieval effectiveness (e.g., nDCG/MAP against a held-out query–relevance setup) and faithfulness (grounding of generated text in the RDF graph / profile) and reports rankings and trade-offs. There is no claimed first-principles derivation, uniqueness theorem, or fitted parameter renamed as a prediction. Unconstrained rewriting’s higher retrieval and lower faithfulness are measured outcomes, not tautologies of the experimental design. Self-citations, if any, are ordinary background and not load-bearing for the central ranking. Faithfulness-metric validity is a correctness/causal-interpretation concern (as the skeptic notes), not circularity under the enumerated kinds. Score 0 is the honest finding: the derivation chain is self-contained experimental comparison against external benchmarks.

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

Empirical IR evaluation paper. Central claims rest on experimental design choices (six generation settings; dual retrieval+faithfulness metrics; RDF dataset corpus) rather than free physical parameters or invented entities. Without full text, metric definitions, prompt designs, and dataset-selection rules cannot be audited in detail; those design choices function as the load-bearing assumptions.

assumptions (3)
  • domain assumption Retrieval effectiveness and faithfulness are jointly sufficient axes to characterize synthetic metadata quality for RDF dataset search.
    Abstract positions the dual evaluation as the right system-level framing; this is a modeling choice about what 'good' synthetic metadata means, not a theorem.
  • domain assumption The six generation settings (simple rewriting through profile-grounded and agentic graph-based) adequately span the relevant design space of LLM metadata generation for RDF.
    Comparative claims about 'unconstrained' vs 'more grounded' settings depend on this coverage assumption; full operationalization is not in the abstract.
  • ad hoc to paper Faithfulness metrics can identify unsupported semantic expansion as distinct from legitimate clarification of true latent dataset properties.
    The causal claim that retrieval gains are driven by unsupported expansion requires that the faithfulness measure cleanly separates invention from true coverage; this is paper-specific and not a standard IR axiom.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Faithful or Findable? Evaluating LLM-Generated Metadata for RDF Dataset Search." pith.science (2026). https://pith.science/paper/6ZLHS73Q

@misc{pith2026260705970,
  author       = {Pith},
  title        = {Pith review of: Faithful or Findable? Evaluating LLM-Generated Metadata for RDF Dataset Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ZLHS73Q}},
  note         = {Machine review of arXiv:2607.05970}
}
read the original abstract

Dataset search depends heavily on metadata, making LLM-generated metadata a consequential form of synthetic content in retrieval systems. We study six metadata-generation settings for RDF datasets, ranging from simple rewriting to profile-grounded and agentic graph-based generation, and evaluate them jointly for retrieval effectiveness and faithfulness. Unconstrained metadata rewriting delivers the strongest retrieval gains over the original metadata, but it is also the least faithful, showing that search improvements can be driven by unsupported semantic expansion. More grounded settings substantially improve faithfulness, and profile-grounded rewriting provides the most balanced trade-off between retrieval effectiveness and grounding. These findings position synthetic metadata as a system-level IR problem in which effectiveness, provenance, and trust must be evaluated together.

Figures

Figures reproduced from arXiv: 2607.05970 by the authors.

Figure 1
Figure 1. Tables 2 and Fig. 1 indicate that [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. Failure modes across the different metadata genera [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [1]

    Sören Auer, Jan Demter, Michael Martin, and Jens Lehmann. 2012. LODStats – An Extensible Framework for High-Performance Dataset Analytics. InKnowledge Engineering and Knowledge Management: 18th International Conference, EKA W

    work page 2012

  2. [2]

    doi:10.1007/978-3-642-33876-2_31

    353–362. doi:10.1007/978-3-642-33876-2_31

    work page doi:10.1007/978-3-642-33876-2_31

  3. [3]

    Alex Bogatu, Alvaro A. A. Fernandes, Norman W. Paton, and Nikolaos Kon- stantinou. 2020. Dataset Discovery in Data Lakes. InProceedings of the 36th IEEE International Conference on Data Engineering (ICDE 2020). 709–720. doi:10.1109/ICDE48307.2020.00067

    work page doi:10.1109/icde48307.2020.00067 2020

  4. [4]

    Veljkovic, Darko Pu- flovic, and Leonid Stoimenov

    Milos Bogdanovic, Milena Frtunic Gligorijevic, Natasa Z. Veljkovic, Darko Pu- flovic, and Leonid Stoimenov. 2023. Cross-Portal Metadata Alignment – Connect- ing Open Data Portals through Means of Formal Concept Analysis.Information Sciences637 (2023), 118958. doi:10.1016/j.ins.2023.118958

    work page doi:10.1016/j.ins.2023.118958 2023

  5. [5]

    Dan Brickley, Matthew Burgess, and Natasha Noy. 2019. Google Dataset Search: Building a Search Engine for Datasets in an Open Web Ecosystem. InThe World Wide Web Conference (WWW ’19). 1365–1375. doi:10.1145/3308558.3313685

    work page doi:10.1145/3308558.3313685 2019

  6. [6]

    Adriane Chapman, Elena Simperl, Laura Koesten, George Konstantinidis, Luis- Daniel Ibáñez-Gonzalez, Emilia Kacprzak, and Paul Groth. 2020. Dataset search: a survey.The VLDB Journal29, 1 (2020), 251–272. doi:10.1007/s00778-019-00564-x

    work page doi:10.1007/s00778-019-00564-x 2020

  7. [7]

    Pan, Yiqun Liu, Emine Yilmaz, Jimmy Lin, Laura Koesten, Mark Sanderson, Frank McCown, Yashar Moshfeghi, and Gong Cheng

    Qiaosheng Chen, Jiaoyan Chen, Xi Fan, Jeff Z. Pan, Yiqun Liu, Emine Yilmaz, Jimmy Lin, Laura Koesten, Mark Sanderson, Frank McCown, Yashar Moshfeghi, and Gong Cheng. 2024. ACORDAR 2.0: A Test Collection for Ad Hoc Dataset Re- trieval with Densely Pooled Datasets and Question-Style Queries. InProceedings of the 47th International ACM SIGIR Conference on Re...

    work page doi:10.1145/3626772.3657866 2024

  8. [8]

    Pan, Yiqun Liu, Emine Yilmaz, Jimmy Lin, Laura Koesten, Mark Sanderson, Frank McCown, Yashar Moshfeghi, and Gong Cheng

    Qiaosheng Chen, Jiaoyan Chen, Xi Fan, Jeff Z. Pan, Yiqun Liu, Emine Yilmaz, Jimmy Lin, Laura Koesten, Mark Sanderson, Frank McCown, Yashar Moshfeghi, and Gong Cheng. 2024. Enhancing Dataset Search with Compact Data Snippets. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’24). 1093–...

    work page doi:10.1145/3626772 2024

Show all 20 references
  1. [9]

    Breslin, Elena Demidova, Stefan Dietze, Julian Szymański, and Konstantin Todorov

    Mohamed Ben Ellefi, Zohra Bellahsene, John G. Breslin, Elena Demidova, Stefan Dietze, Julian Szymański, and Konstantin Todorov. 2018. RDF Dataset Profiling – A Survey of Features, Methods, Vocabularies and Applications.Semantic Web9, 5 (2018), 677–705. doi:10.3233/SW-180294

    2018 doi

  2. [10]

    Ivan Ermilov, Diego Kontokostas, Sören Auer, Jens Lehmann, and Axel- Cyrille Ngonga Ngomo. 2016. LODStats: The Data Web Census Dataset. In The Semantic Web – ISWC 2016: 15th International Semantic Web Conference, Part II. 38–46. doi:10.1007/978-3-319-46547-0_5

    2016 doi

  3. [11]

    Hideaki Honma, Shoji Noguchi, Makx Dekkers, Makoto Murata, and Takahiro Kawamura. 2014. Extracting Description Set Profiles from RDF Datasets using Metadata Instances and SPARQL Queries. InProceedings of the International Conference on Dublin Core and Metadata Applications 201...

    2014

  4. [12]

    Petr Křemen and Martin Nečaský. 2019. Improving Discoverability of Open Government Data with Rich Metadata Descriptions Using Semantic Government Vocabulary.Journal of Web Semantics55 (2019), 1–20. doi:10.1016/j.websem.2018. 12.009

    2019 doi

  5. [13]

    Sylvain Kübler, Jérémy Robert, Sebastian Neumaier, Jürgen Umbrich, and Yves Le Traon. 2018. Comparison of Metadata Quality in Open Data Portals Using the Analytic Hierarchy Process.Government Information Quarterly35, 1 (2018), 13–29. doi:10.1016/j.giq.2017.11.003

    2018 doi

  6. [14]

    Weizhi Lin, Yashar Moshfeghi, Frank McCown, Laura Koesten, Mark Sanderson, Jinfeng Zhang, Yiqun Liu, Emine Yilmaz, and Jimmy Lin. 2022. ACORDAR: A Test Collection for Ad Hoc Content-Based (RDF) Dataset Retrieval. InProceedings of the 45th International ACM SIGIR Conference on ...

    2022 doi

  7. [15]

    Sebastian Neumaier, Jürgen Umbrich, and Axel Polleres. 2016. Automated Quality Assessment of Metadata across Open Data Portals.ACM Journal of Data and Information Quality8, 1 (2016), 2:1–2:29. doi:10.1145/2964909

    2016 doi

  8. [16]

    Zarazaga-Soria

    Javier Nogueras-Iso, Sergio Lacasta, Núria Minguillón Alfonso, Dhiraj Khatiwada, and Francisco J. Zarazaga-Soria. 2021. Quality of Metadata in Open Data Portals. IEEE Access9 (2021), 60364–60382. doi:10.1109/ACCESS.2021.3073455

    2021 doi

  9. [17]

    Paton, Jiaoyan Chen, and Zhenyu Wu

    Norman W. Paton, Jiaoyan Chen, and Zhenyu Wu. 2024. Dataset Discovery and Exploration: A Survey.Comput. Surveys56, 4 (2024), 102:1–102:37. doi:10.1145/ 3626521

    2024

  10. [18]

    Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Frame- work: BM25 and Beyond.Foundations and Trends in Information Retrieval3, 4 (2009), 333–389. doi:10.1561/1500000019

    2009 doi

  11. [19]

    Haoxiang Zhang, Yurong Liu, Wei-Lun Hung, Aécio S. R. Santos, and Juliana Freire. 2025. AutoDDG: Automated Dataset Description Generation using Large Language Models. arXiv preprint. arXiv:2502.01050 doi:10.48550/arXiv.2502. 01050

    2025 doi

  12. [20]

    Xiao Zhou, Qiaosheng Chen, Jiageng Chen, and Gong Cheng. 2025. 𝜇DS: Multi- Objective Data Snippet Extraction for Dataset Search. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’25). 116–126. doi:10.1145/...

    2025 doi

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.