Pith. sign in

REVIEW 3 major objections 4 minor 36 references

For finding the source paper behind a social-media claim, verification re-ranking beats similarity alone, reaching an MRR@5 of 0.758.

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

2026-08-01 22:04 UTC pith:PKIM3KH3

load-bearing objection Solid comparative study, but the headline MRR@5 is an in-sample selection artifact; the verification-based gain is provisional until a held-out test run exists. the 3 major comments →

arxiv 2607.15875 v1 pith:PKIM3KH3 submitted 2026-07-17 cs.IR

Scientific Claim-Source Retrieval Revisited: A Comparative Study of Style Transfer and Re-Ranking

Tobias Schreieder , Harsh Khandelwal , Yu-Ling Zhong , Michael F\"arber This is my paper
classification cs.IR
keywords scientific claim-source retrievalre-rankingtext style transfermultilingual retrievalverification reasoningpublication metadatalarge language modelsMRR
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.

This paper asks how to retrieve the scientific publication that a social-media claim is based on, when the claim is written in colloquial language, possibly in a different language, and refers to the source only implicitly. It compares multilingual representations, style transfer, and two generations of re-ranking, and reports that a verification-based re-ranker—one that reasons over whether each candidate source actually supports the claim—outperforms semantic similarity alone. The best pipeline reaches MRR@5 of 0.758, a +0.025 gain over the semantic baseline. The paper argues that the effect is specific: translating claims to English helps, metadata helps mostly sparse retrieval, style transfer helps only some models, and entity overlap actually hurts.

Core claim

The central discovery is that verification reasoning can refine a ranking that semantic similarity has already produced. On the benchmark's validation set, the authors build a pipeline stepwise: translate non-English claims to English, append publication metadata to source representations, retrieve top-100 candidates with a strong dense retriever, re-rank with a large generative cross-encoder, and finally apply a zero-shot, verification-based re-ranker that prompts a reasoning LLM to select the single candidate source that truly underlies the claim. That final step lifts MRR@5 from 0.733 to 0.758. The verification prompt explicitly assesses source verification, evidence alignment, contextual

What carries the argument

Verification-based re-ranking: a zero-shot prompted LLM is given the claim and the top-10 candidates and must pick the one source that best satisfies 'does this paper actually back the claim?' using checks for source verification, evidence alignment, contextual specificity, and conflicting findings. The chosen source is given top rank. This is the mechanism behind the paper's best result; it is contrasted with similarity-based re-ranking (cross-encoders that score claim–source pairs semantically) and two other signal-based re-rankers (attribution/atomic-fact support rate and entity overlap proportion).

Load-bearing premise

The reported 0.758 MRR@5 is measured on the same 4,993-claim validation set that was used to choose each pipeline stage; if that set is not representative of unseen claims, the measured advantage of verification-based re-ranking may shrink or disappear.

What would settle it

Hold out a random sample of claims (or use the official test split) from the benchmark; compute MRR@5 for the verification pipeline versus the semantic-only baseline (retrieval plus cross-encoder re-ranking). If the verification pipeline does not significantly beat the baseline on the held-out set (paired test, p < 0.05), the paper's central claim fails.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Fact-checking pipelines that already use similarity re-ranking can gain additional accuracy by adding an LLM verification step on the top candidates.
  • For non-English claims, translating to English before retrieval is beneficial even for strong dense retrievers (e.g., +0.167 for one dense model).
  • Publication metadata is a cheap way to capture implicit references, most useful for lexical (sparse) models.
  • Style transfer's benefit is conditional: formal style helps the sparse model, question style helps dense models, and the strongest retriever prefers original claims.
  • Re-ranker scale matters: the largest generative re-ranker improves MRR@5 by +0.100, while smaller re-rankers can degrade performance.

Where Pith is reading between the lines

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

  • I infer that the 0.758 figure, computed on the same validation set used to select each pipeline stage, is likely an optimistic estimate; a held-out test set would probably shrink the verification gain, though it may persist.
  • The verification prompt structure may transfer to other 'claim-to-document' tasks, such as health advisories or policy briefs, where the source is referenced implicitly.
  • Entity overlap hurting despite intuition suggests a learned or weighted entity signal, rather than raw proportion, might revive entity-based re-ranking.
  • A practical cost-accuracy trade-off could be explored by applying verification to top-5 or top-3 candidates instead of top-10, potentially retaining most of the gain at lower LLM cost.

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 / 4 minor

Summary. The paper presents a comparative study of scientific claim–source retrieval on the CheckThat! 2026 benchmark, covering multilingual claim handling, publication metadata, style transfer, similarity-based re-ranking, and three proposed signal-based re-rankers. Using validation data only (4,993 claims), the authors report that translated claims outperform original and bilingual representations, metadata helps especially BM25, style transfer helps some retrieval models, similarity-based re-ranking with Qwen3-8B improves MRR@5 from 0.633 to 0.733, and verification-based re-ranking with Kimi-2.6 further improves it to 0.758. The paper releases code and resources.

Significance. If the headline results hold out-of-sample, the paper would provide useful evidence for practitioners: translation is more effective than bilingual encoding, metadata is valuable for sparse retrieval, question-style claims help certain dense retrievers, and a large LLM can provide a small but potentially real verification signal beyond cross-encoder similarity. The study is broad in scope and the reported tables are internally consistent. Strengths include the use of a standard benchmark, evaluation of multiple model families, repeated runs with reported stability, and public code. However, the central numeric claim is currently an in-sample, sequentially selected statistic, so the generalization of the verification-based gain is not established.

major comments (3)
  1. [§3.1, §5 (intro), Table 5] All experiments are on the validation set, and §5 states that each stage 'builds upon the strongest results of the previous one' on that same set. The final Kimi-2.6 verification result (0.758) is a maximum selected over 3 reasoning LLMs × 3 signal types, added to previously selected stages, all evaluated on the same 4,993 claims. The paired t-test in Table 5 is also computed on this same set and does not account for model selection. The +0.025 improvement over the Qwen3-8B baseline is therefore an in-sample selected estimate, not evidence of generalization. A held-out test set or nested/out-of-sample selection procedure is needed to support the abstract's claim.
  2. [§5 intro, §5.3, §5.4] The paper is framed as a comparative study of style transfer and re-ranking, but the two are not combined. The re-ranking stage (Tables 4 and 5) uses GritLM retrieval on original, untransformed claims, even though Table 3 shows style transfer improves several models (BM25 formal: +0.019; E5 question: +0.026). No experiment feeds the best style-transferred claims into the re-ranking pipeline. Thus the paper cannot tell whether style transfer and re-ranking are complementary or redundant, and the final 'best overall' pipeline omits the style-transfer gains that the paper itself reports for other retrieval models.
  3. [Abstract, §5.3, Table 3] The abstract claims that style transfer approaches 'improve retrieval performance for most models.' Table 3 does not fully support this generalization: for GritLM every style reduces MRR@5 (0.633 baseline vs. 0.576–0.630), BM25 Abstract drops to 0.298, and only a minority of configurations are marked as statistically significant. The claim should be qualified by model type and style, and the significance tests should be corrected for the number of comparisons.
minor comments (4)
  1. [§3.1] The decision to restrict to validation data is stated, but the existence and availability of a test split is not discussed. If CheckThat! 2026 provides a test set, using it (or explaining why it is unavailable) would substantially strengthen the paper.
  2. [§5.3] The text reports standard deviations '≤0.004' but Table 3 does not show them. Since LLM generation is stochastic, including the standard deviations in the table (or an appendix) would make the stability claim checkable.
  3. [§4.4] The listwise re-ranker Jina and the generative re-rankers Qwen3 are described briefly; a sentence on the input length / context window used for the 100 candidates would help reproducibility, since listwise re-rankers often truncate the candidate list.
  4. [§4.5] The attribution and entity signals are aggregated as simple proportions, but the prompts are not included in the paper. Given the 'zero-shot prompt' design, releasing the exact prompt templates (in addition to code) is important for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the reported gains are in-paper empirical evaluations of pretrained models and zero-shot prompts; self-citations are contextual and no result reduces to a fit or to a self-citation chain.

full rationale

The paper's central claims are empirical comparisons computed directly on the public CheckThat! 2026 validation set using off-the-shelf pretrained retrieval models and zero-shot LLM prompts. No parameter is fitted to the evaluation labels, and no reported score is defined as the output of a fitted model. The headline verification-based re-ranking gain (0.733 to 0.758, Table 5) is obtained by applying a zero-shot prompt that selects a top candidate and re-scores it; this is an LLM prediction, not a construction equivalent to the retrieval target. The reuse of the authors' own prompt templates from [18] (Section 4.3) is an explicit methodological choice for comparability, and the citation to [18] for the observation that formalizing claims improves sparse retrieval is contextual; the paper re-evaluates style transfer on this dataset in Table 3. Similarly, the decision to keep source documents unchanged relies on prior work but is an experimental-design choice, not the derivation of the central result. The main non-circularity concern is the absence of a held-out test set: Section 3.1 states 'Since our study does not involve training, we restrict our experiments to the validation set,' and Section 5's sequential strategy ('each stage builds upon the strongest results of the previous one') means the reported MRR@5 is an in-sample, stage-wise selected estimate, with paired t-tests computed on the same set. This is a generalization/selection-bias risk, not circularity, and therefore does not raise the circularity score beyond a minor allowance for the cited prior work.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper does not introduce mathematical free parameters; the listed design choices are hand-selected thresholds that shape the results. The central claim rests on the quality of the shared-task labels, the validity of MRR@5 as the metric, and the meaning-preservation of LLM transformations; none of these are independently verified in the paper. No new entities are introduced.

free parameters (3)
  • similarity re-ranking candidate pool size = 100
    Section 4.4: re-rankers operate on the top-100 GritLM results; correct sources outside this pool cannot be recovered, setting a ceiling on re-ranking gains.
  • signal-based re-ranking candidate pool size = 10
    Section 4.5: LLM reasoning is limited to the top-10 candidates for cost; the reported MRR@5 improvements are contingent on this pool.
  • author metadata truncation = first and last three authors
    Section 4.2: 'we include only the first and last three authors' to limit noise; this changes the lexical cues available to retrieval models.
axioms (4)
  • domain assumption The CheckThat! 2026 validation set labels are correct and complete, with the true source for each claim present in the 10,000-publication candidate set.
    All conclusions depend on dataset quality; Section 3.1 describes the dataset but provides no label-quality or completeness analysis.
  • domain assumption MRR@5 on the validation set is a valid measure of scientific claim-source retrieval performance.
    The paper uses MRR@5 exclusively (Section 5) without reporting recall or test-set performance; if the metric or split is unrepresentative, the ranking of approaches may not transfer.
  • domain assumption Zero-shot LLM translation and style transfer preserve the factual content of claims.
    Translation and style transfer (Sections 4.1 and 4.3) are not human-evaluated; retrieval gains could stem from output artifacts rather than improved alignment.
  • domain assumption LLM verification-based selection identifies genuine source support rather than superficial similarity.
    The verification re-ranker in Section 4.5 selects a single candidate based on source verification, evidence alignment, contextual specificity, and exclusion checks, but there is no external verification that these judgements are factually correct.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Scientific Claim-Source Retrieval Revisited: A Comparative Study of Style Transfer and Re-Ranking." pith.science (2026). https://pith.science/paper/PKIM3KH3

@misc{pith2026260715875,
  author       = {Pith},
  title        = {Pith review of: Scientific Claim-Source Retrieval Revisited: A Comparative Study of Style Transfer and Re-Ranking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PKIM3KH3}},
  note         = {Machine review of arXiv:2607.15875}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Scientific claims shared on social media are often difficult to verify and may contribute to the spread of misinformation. To address this challenge, automated fact verification systems require scientific claim-source retrieval, the task of identifying the source publication underlying a given claim. However, claims often differ substantially from their source publications in language, style, and specificity, making retrieval challenging. We present a comparative study of scientific claim-source retrieval on the CheckThat! 2026 benchmark across sparse and dense retrieval models. Our results show that translating claims into English outperforms both original and bilingual claim representations, while incorporating publication metadata provides additional retrieval gains by capturing indirect source references. In addition, we analyze four style transfer approaches and find that they improve retrieval performance for most models, although the optimal style depends on the underlying retrieval objective. Finally, we investigate similarity- and signal-based re-ranking approaches, introducing three novel re-ranking models based on attribution, entity overlap, and verification-based reasoning. Verification-based re-ranking yields additional gains beyond semantic similarity and achieves the best overall performance with an MRR@5 of 0.758.

Figures

Figures reproduced from arXiv: 2607.15875 by Harsh Khandelwal, Michael F\"arber, Tobias Schreieder, Yu-Ling Zhong.

Figure 1
Figure 1. Figure 1: Illustration of the style transfer process. Given a claim, the LLM generates stylistic variants to create diverse representations while preserving the original meaning. 4.3 Style Transfer for Scientific Claims As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the signal-based re-ranking process. Signals extracted from a claim are used to re-rank top-k sources, where sources are scored according to signal matching. 4.4 Similarity-based Re-Ranking Re-ranking has proven to be an effective component in scientific claim–source retrieval, as top-performing systems in the CheckThat! 2025 shared task incor￾porated re-ranking approaches to improve retrieval … 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

36 extracted references · 6 canonical work pages

  1. [1]

    In: Carrillo-de Albornoz, J., García Seco de Herrera, A., Gonzalo, J., Plaza, L., Mothe, J., Piroi, F., Rosso, P., Spina, D., Faggioli, G., Ferro, N

    Alam, F., Struß, J.M., Chakraborty, T., Dietze, S., Hafid, S., Korre, K., Muti, A., Nakov, P., Ruggeri, F., Schellhammer, S., Setty, V., Sundriyal, M., Todorov, K., Venktesh, V.: Overview of the clef-2025 checkthat! lab: Subjectivity, fact-checking, claim normalization, and retrieval. In: Carrillo-de Albornoz, J., García Seco de Herrera, A., Gonzalo, J., ...

    2025

  2. [2]

    In: Proceedings of the 34th ACM International Conference on Information and Knowledge Management

    Besrour, I., He, J., Schreieder, T., Färber, M.: Squai: Scientific question-answering with multi-agent retrieval-augmented generation. In: Proceedings of the 34th ACM International Conference on Information and Knowledge Management. p. 6603–6608. CIKM ’25, Association for Computing Machinery, New York, NY, USA (2025).https://doi.org/10.1145/3746252.3761471

    arXiv 2025

  3. [3]

    In: Ku, L.W., Martins, A., Srikumar, V

    Chen, J., Xiao, S., Zhang, P., Luo, K., Lian, D., Liu, Z.: M3-embedding: Multi- linguality, multi-functionality, multi-granularity text embeddings through self- knowledge distillation. In: Ku, L.W., Martins, A., Srikumar, V. (eds.) Findings of the Association for Computational Linguistics: ACL 2024. pp. 2318–2335. As- sociation for Computational Linguisti...

    doi:10.18653/v1/2024.findings-acl.137 2024

  4. [4]

    In: Duh, K., Gomez, H., Bethard, S

    Chen, J., Kim, G., Sriram, A., Durrett, G., Choi, E.: Complex claim verification with evidence retrieved in the wild. In: Duh, K., Gomez, H., Bethard, S. (eds.) Proceedings of the 2024 Conference of the North American Chapter of the Asso- ciation for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). pp. 3569–3587. Association...

    doi:10.18653/v1/2024.naacl-long.196 2024

  5. [5]

    In: Schlichtkrull, M., Chen, Y., Whitehouse, C., Deng, Z., Akhtar, M., Aly, R., Guo, Z., Christodoulopou- los, C., Cocarascu, O., Mittal, A., Thorne, J., Vlachos, A

    Churina, S., Barik, A.M., Phaye, S.R.: Improving evidence retrieval on claim verification pipeline through question enrichment. In: Schlichtkrull, M., Chen, Y., Whitehouse, C., Deng, Z., Akhtar, M., Aly, R., Guo, Z., Christodoulopou- los, C., Cocarascu, O., Mittal, A., Thorne, J., Vlachos, A. (eds.) Proceedings of the Seventh Fact Extraction and VERificat...

  6. [6]

    Public Understanding of Science32(7), 860–869 (2023).https://doi.org/10.1177/09636625231166552

    Guenther, L., Wilhelm, C., Oschatz, C., Brück, J.: Science communication on twitter: Measuring indicators of engagement and their links to user interaction in communication scholars’ tweet content. Public Understanding of Science32(7), 860–869 (2023).https://doi.org/10.1177/09636625231166552

    doi:10.1177/09636625231166552 2023

  7. [7]

    Quantitative Science Studies2(4), 1486–1510 (2021).https: //doi.org/10.1162/qss_a_00168

    Haunschild, R., Bornmann, L., Potnis, D., Tahamtan, I.: Investigating dissem- ination of scientific information on twitter: A study of topic networks in opi- oid publications. Quantitative Science Studies2(4), 1486–1510 (2021).https: //doi.org/10.1162/qss_a_00168

    work page doi:10.1162/qss_a_00168 2021

  8. [8]

    In: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Liao, H., Peng, J., Huang, Z., Zhang, W., Li, G., Shu, K., Xie, X.: Muser: A multi-step evidence retrieval enhancement framework for fake news detection. In: Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. p. 4461–4472. KDD ’23, Association for Computing Machinery, New York, NY, USA (2023).https://doi.org/10.1145/3580...

    arXiv 2023

  9. [9]

    In: Hacid, H., Cellary, W., Wang, H., Paik, H.Y., Zhou, R

    Ma, W., Chao, W., Luo, Z., Jiang, X.: Claim retrieval in twitter. In: Hacid, H., Cellary, W., Wang, H., Paik, H.Y., Zhou, R. (eds.) Web Information Systems En- gineering – WISE 2018. pp. 297–307. Springer International Publishing, Cham (2018).https://doi.org/10.1007/978-3-030-02922-7_20 14 T. Schreieder et al

    work page doi:10.1007/978-3-030-02922-7_20 2018

  10. [10]

    In: Bouamor, H., Pino, J., Bali, K

    Min,S.,Krishna,K.,Lyu,X.,Lewis,M.,Yih,W.t.,Koh,P.,Iyyer,M.,Zettlemoyer, L., Hajishirzi, H.: FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In: Bouamor, H., Pino, J., Bali, K. (eds.) Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 12076–12100. Association for Computat...

    doi:10.18653/v1/2023.emnlp-main.741 2023

  11. [11]

    In: The Thirteenth International Conference on Learning Representations (2025),https://openreview.net/forum? id=BC4lIvfSzv

    Muennighoff, N., SU, H., Wang, L., Yang, N., Wei, F., Yu, T., Singh, A., Kiela, D.: Generative representational instruction tuning. In: The Thirteenth International Conference on Learning Representations (2025),https://openreview.net/forum? id=BC4lIvfSzv

    2025

  12. [12]

    In:Goldberg,Y.,Kozareva,Z.,Zhang,Y.(eds.)Proceedingsofthe2022Conference on Empirical Methods in Natural Language Processing

    Ni, J., Qu, C., Lu, J., Dai, Z., Hernandez Abrego, G., Ma, J., Zhao, V., Luan, Y., Hall, K., Chang, M.W., Yang, Y.: Large dual encoders are generalizable retrievers. In:Goldberg,Y.,Kozareva,Z.,Zhang,Y.(eds.)Proceedingsofthe2022Conference on Empirical Methods in Natural Language Processing. pp. 9844–9855. Association for Computational Linguistics, Abu Dhab...

    doi:10.18653/v1/2022.emnlp-main.669 2022

  13. [13]

    In: SIGIR ’94

    Robertson, S.E., Walker, S.: Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval. In: SIGIR ’94. p. 232–241. Springer London, London (1994).https://doi.org/10.5555/188490.188561

    arXiv 1994

  14. [14]

    In: Harman, D.K

    Robertson, S.E., Walker, S., Jones, S., Hancock-Beaulieu, M., Gatford, M.: Okapi at TREC-3. In: Harman, D.K. (ed.) Proceedings of The Third Text RE- trieval Conference, TREC 1994, Gaithersburg, Maryland, USA, November 2-4,

    1994

  15. [15]

    In: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum

    Sager,P.J.,Kamaraj,A.,Grewe,B.F.,Stadelmann,T.:Deepretrievalatcheckthat! 2025: Identifying scientific papers from implicit social media mentions via hybrid retrieval and re-ranking. In: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum. CLEF 2025, Madrid, Spain (2025),https://ceur-ws. org/Vol-4038/paper_89.pdf

    2025

  16. [16]

    In: Sil, A., Lin, X.V

    Samarinas, C., Hsu, W., Lee, M.L.: Improving evidence retrieval for automated explainable fact-checking. In: Sil, A., Lin, X.V. (eds.) Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Demonstrations. pp. 84–91. Associ- ation for Computational Linguistics, Online...

    2021

  17. [17]

    In: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum

    Schofield, J., Tian, S., Truong, H.T.T., Heil, M.: Ds@gt at checkthat! 2025: Ex- ploring retrieval and reranking pipelines for scientific claim source retrieval on social media discourse. In: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum. CLEF 2025, Madrid, Spain (2025),https://ceur-ws.org/ Vol-4038/paper_93.pdf

    2025

  18. [18]

    In: Working Notes of CLEF 2025 - Con- ference and Labs of the Evaluation Forum

    Schreieder, T., Färber, M.: Claim2source at checkthat! 2025: Zero-shot style trans- fer for scientific claim-source retrieval. In: Working Notes of CLEF 2025 - Con- ference and Labs of the Evaluation Forum. CLEF 2025, Madrid, Spain (2025), https://ceur-ws.org/Vol-4038/paper_94.pdf

    2025

  19. [19]

    In: Liakata, M., Moreira, V.P., Zhang, J., Jurgens, D

    Schreieder, T., Schopf, T., Färber, M.: Attribution, citation, and quotation: A survey of evidence-based text generation with large language models. In: Liakata, M., Moreira, V.P., Zhang, J., Jurgens, D. (eds.) Proceedings of the 64th Annual MeetingoftheAssociationforComputationalLinguistics(Volume1:LongPapers). Scientific Claim–Source Retrieval Revisited...

    work page doi:10.18653/v1/2026.acl-lo 2026

  20. [20]

    In: Shi, W., Yu, W., Asai, A., Jiang, M., Durrett, G., Hajishirzi, H., Zettle- moyer, L

    Shavarani, H., Sarkar, A.: Entity retrieval for answering entity-centric questions. In: Shi, W., Yu, W., Asai, A., Jiang, M., Durrett, G., Hajishirzi, H., Zettle- moyer, L. (eds.) Proceedings of the 4th International Workshop on Knowledge- Augmented Methods for Natural Language Processing. pp. 1–17. Association for Computational Linguistics, Albuquerque, ...

    work page doi:10.18653/v1/2025.knowledgenlp-1.1 2025

  21. [21]

    In: Jose, J.M., Yilmaz, E., Magalhães, J., Castells, P., Ferro, N., Silva, M.J., Martins, F

    Soleimani, A., Monz, C., Worring, M.: Bert for evidence retrieval and claim verification. In: Jose, J.M., Yilmaz, E., Magalhães, J., Castells, P., Ferro, N., Silva, M.J., Martins, F. (eds.) Advances in Information Retrieval. pp. 359–

  22. [22]

    In: Schlichtkrull, M., Chen, Y., Whitehouse, C., Deng, Z., Akhtar, M., Aly, R., Guo, Z., Christodoulopoulos, C., Cocarascu, O., Mittal, A., Thorne, J., Vlachos, A

    Sriram, A., Xu, F., Choi, E., Durrett, G.: Contrastive learning to improve retrieval for real-world fact checking. In: Schlichtkrull, M., Chen, Y., Whitehouse, C., Deng, Z., Akhtar, M., Aly, R., Guo, Z., Christodoulopoulos, C., Cocarascu, O., Mittal, A., Thorne, J., Vlachos, A. (eds.) Proceedings of the Seventh Fact Extraction and VERification Workshop (F...

    2024

  23. [23]

    In: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum

    Staudinger, M., El-Ebshihy, A., Kusa, W., Piroi, F., Hanbury, A.: Atom at check- that! 2025: Retrieve the implicit - scientific evidence retrieval. In: Working Notes of CLEF 2025 - Conference and Labs of the Evaluation Forum. CLEF 2025, Madrid, Spain (2025),https://ceur-ws.org/Vol-4038/paper_98.pdf

    2025

  24. [24]

    In: Campos, R., Jatowt, A., Lan, Y., Aliannejadi, M., Bauer, C., MacAvaney, S., Anand, A., Ren, Z., Ver- berne, S., Bai, N., Mansoury, M

    Struß,J.M.,Schellhammer,S.,Dietze,S.,V.,V.,Setty,V.,Chakraborty,T.,Nakov, P., Anand, A., Chungkham, P., Hafid, S., Sahnan, D., Todorov, K.: The clef-2026 checkthat! lab: Advancing multilingual fact-checking. In: Campos, R., Jatowt, A., Lan, Y., Aliannejadi, M., Bauer, C., MacAvaney, S., Anand, A., Ren, Z., Ver- berne, S., Bai, N., Mansoury, M. (eds.) Adva...

    2026

  25. [25]

    Suarez-Lledo, V., Alvarez-Galvez, J.: Prevalence of health misinformation on social media: Systematic review. J. Med. Internet Res.23(1), e17187 (Jan 2021).https: //doi.org/10.2196/17187

    doi:10.2196/17187 2021

  26. [26]

    In: Proceedings of the Workshop on Combating Online Hostile Posts in Regional Languages during Emergency Situations

    Sundriyal, M., Malhotra, G., Akhtar, M.S., Sengupta, S., Fano, A., Chakraborty, T.: Document retrieval and claim verification to mitigate COVID-19 misinfor- mation. In: Proceedings of the Workshop on Combating Online Hostile Posts in Regional Languages during Emergency Situations. pp. 66–74. Association for Computational Linguistics, Dublin, Ireland (2022...

    2022

  27. [27]

    Annual Review of Public Health41(Volume 41, 2020), 433–451 (2020).https://doi.org/10.1146/ annurev-publhealth-040119-094127

    Swire-Thompson, B., Lazer, D.: Public health and online misinfor- mation: Challenges and recommendations. Annual Review of Public Health41(Volume 41, 2020), 433–451 (2020).https://doi.org/10.1146/ annurev-publhealth-040119-094127

    2020

  28. [28]

    Discover Computing28(1), 27 (2025).https://doi.org/10.1007/s10791-025-09505-5

    Upadhyay, R., Viviani, M.: Enhancing health information retrieval with RAG by prioritizing topical relevance and factual accuracy. Discover Computing28(1), 27 (2025).https://doi.org/10.1007/s10791-025-09505-5

    work page doi:10.1007/s10791-025-09505-5 2025

  29. [29]

    Schreieder et al

    Wang,F.,Li,Y.,Xiao,H.:jina-reranker-v3:Lastbutnotlateinteractionforlistwise document reranking (2025),https://arxiv.org/abs/2509.25085 16 T. Schreieder et al

    arXiv 2025

  30. [30]

    Wang, L., Yang, N., Huang, X., Jiao, B., Yang, L., Jiang, D., Majumder, R., Wei, F.: Text embeddings by weakly-supervised contrastive pre-training (2024), https://arxiv.org/abs/2212.03533

    Pith/arXiv arXiv 2024

  31. [31]

    In: Bouamor, H., Pino, J., Bali, K

    Zhang, H., Zhang, R., Guo, J., de Rijke, M., Fan, Y., Cheng, X.: From relevance to utility: Evidence retrieval with feedback for fact verification. In: Bouamor, H., Pino, J., Bali, K. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 6373–6384. Association for Computational Linguistics, Singa- pore (Dec 2023).https://doi.or...

    doi:10.18653/v1/2023.findings-emnlp.422 2023

  32. [32]

    Zhang, Y., Li, M., Long, D., Zhang, X., Lin, H., Yang, B., Xie, P., Yang, A., Liu, D., Lin, J., Huang, F., Zhou, J.: Qwen3 embedding: Advancing text embedding and reranking through foundation models (2025),https://arxiv.org/abs/2506. 05176

    2025

  33. [33]

    In: Yang, X., Suhartanto, H., Wang, G., Wang, B., Jiang, J., Li, B., Zhu, H., Cui, N

    Zuo, C., Wang, C., Banerjee, R.: Cross-genre retrieval for information integrity: A covid-19 case study. In: Yang, X., Suhartanto, H., Wang, G., Wang, B., Jiang, J., Li, B., Zhu, H., Cui, N. (eds.) Advanced Data Mining and Applications. pp. 495–509. Springer Nature Switzerland, Cham (2023).https://doi.org/10.1007/ 978-3-031-46677-9_34

    2023

  34. [70]

    https://doi.org/10.18653/v1/2024.fever-1.6

    Association for Computational Linguistics, Miami, Florida, USA (Nov 2024). https://doi.org/10.18653/v1/2024.fever-1.6

    work page doi:10.18653/v1/2024.fever-1.6 2024

  35. [366]

    Springer International Publishing, Cham (2020).https://doi.org/10.1007/ 978-3-030-45442-5_45

    2020

  36. [1994]

    500-225, pp

    NIST Special Publication, vol. 500-225, pp. 109–126. National Institute of Standards and Technology (NIST) (1994),http://trec.nist.gov/pubs/trec3/ papers/city.ps.gz

    1994

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.