Pith. sign in

REVIEW 4 major objections 6 minor 39 references

FlashCheck: Exploration of Efficient Evidence Retrieval for Fast Fact-Checking

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Pruning Wikipedia to factual sentences and compressing the neural index speeds up fact-checking by 10x on CPUs and over 20x on GPUs, with only minor loss in verification F1.

desk verdict Solid CPU speedups and a useful dataset, but the GPU >20x headline rests on a CPU baseline and needs re-measurement or qualification. read the letter →

arxiv 2502.05803 v2 pith:QSDAENTI submitted 2025-02-09 cs.IR

Kevin Nanekhan , Venktesh V , Erik Martin , Henrik Vatndal , Vinay Setty , Avishek Anand This is my paper · ORCID
classification cs.IR
keywords fact-checkingevidenceretrievalindexcompressionproductquantizationdenseWikipediavectorlatencyspeedup
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

The paper argues that the retrieval stage of automated fact-checking is the main scalability bottleneck and shows that it can be made over an order of magnitude faster. Its central claim is that by indexing only sentences likely to be verifiable facts—selected by a claim-detection model, by citation presence, or by both—and then compressing the dense vector index with joint product quantization, a fact-checking pipeline runs up to 10.0x faster on CPUs and more than 20.0x faster on GPUs than a classical sparse retrieval plus reranking pipeline over full Wikipedia. The paper reports that this comes with only a minor loss in downstream weighted F1. A live test on claims from the 2024 presidential debate runs 3.4x faster than the classical pipeline with essentially unchanged accuracy. If these results hold beyond the two benchmarks studied, they make real-time and low-resource fact-checking substantially more practical.

What carries the argument

The load-bearing mechanism is the fusion of corpus pruning with joint product quantization. Pruning works by keeping only sentences that a claim-detection model flags as check-worthy or that carry Wikipedia citations (or both), producing a 'fact index' that is about half the size of full Wikipedia and already segmented into evidence-like snippets, so the pipeline's separate sentence-selection stage can be dropped. The compression side uses JPQ, a product-quantization method that replaces each document embedding with a short concatenation of centroid sub-vectors while jointly training the query encoder against the quantized index using a ranking loss over hard negatives; this is what turns dense retrieval from a brute-force scan into fast table lookups. JPQ delivers a compression ratio of $4D/M$ and a speedup ratio of $(D + \log n)/(M + \log n)$, where $D$ is embedding dimensionality and $M$ is the number of codebooks. Together the two techniques produce the reported speedups with weighted F1 loss of only a few points.

What would settle it

Measure evidence recall@k on the Fusion-pruned index against the gold evidence annotations in the two benchmarks; if a substantial share of gold passages are pruned away while end-to-end F1 stays flat, the pruning criterion is fitted to the benchmarks' evidence style, and the speedup would not transfer to claim sets whose evidence lies in uncited context sentences.

Watch

Extended reading notes

Core claim

In the paper's own terms, the discovery is that existing fact-checking pipelines waste most of their retrieval effort on Wikipedia sentences that are not needed for verification. The authors propose three ways to prune the corpus to factual statements—claim-detection-based extraction, citation-based extraction, and a fusion of the two—and show that the pruned corpus cuts disk size and sentence count by roughly half while preserving task accuracy. They then show that applying JPQ index compression to the dense embeddings of the pruned collection shrinks the index by 93% and, because JPQ jointly trains the query encoder and the quantized index, retrieval can actually match or slightly beat uncompressed dense retrieval. The combination of pruning and compression removes the need for a separate sentence-selection stage, and the reported end-to-end latency on CPUs falls from hundreds of milliseconds to tens of milliseconds. The paper states the result as up to 10.0x CPU speedup and more than 20.0x GPU speedup over the classical pipeline.

Load-bearing premise

The load-bearing premise is that a Wikipedia corpus pruned to claim-detectable and citation-backed sentences still contains essentially all the evidence needed to verify claims, so discarding the rest costs little accuracy.

Editorial extensions

If this is right

  • Dropping the sentence-selection stage from the pipeline shortens end-to-end latency beyond the retrieval gain, because the pruned index already returns evidence-sized snippets.
  • The 93% index-size reduction (from about 9.70 GiB to 672.95 MiB on one collection) makes dense retrieval practical on CPU-only machines and memory-limited devices.
  • End-to-end CPU latency on one benchmark drops from 659 ms for sparse retrieve-plus-rerank to 51–63 ms with the compressed pruned index.
  • Live fact-checking of the 2024 presidential debate runs 3.4x faster than the classical pipeline over full Wikipedia with no significant change in weighted F1.
  • Jointly trained compression can slightly outperform uncompressed dense retrieval, because the end-to-end negative sampling improves the query encoder.

Reading between the lines

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

  • Inference: The pruning assumption is likely to transfer only to claim sets whose evidence is citation-backed; claims about emerging events, minor entities, or context-dependent statements would see larger recall drops than these benchmarks show, so practitioners should measure evidence recall on their own claim distribution.
  • Inference: The same prune-and-quantize recipe could be applied to other knowledge sources such as news archives or scientific literature; the main accuracy lever would then be the choice of claim detector rather than the compression step.
  • Inference: The reported speedups imply that dense retrieval over billion-scale corpora on commodity hardware is feasible, which would also benefit open-domain question answering and retrieval-augmented generation, not just fact-checking.
  • Inference: A more direct test of the pruning assumption would be to report evidence recall@k on the pruned index against the gold evidence annotations, isolating whether the small F1 drops come from pruning misses or from quantization noise.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes making fact-checking evidence retrieval faster by pruning a large Wikipedia corpus down to 'factual' sentences (via ClaimBuster-style claim detection, citation extraction, or their fusion) and then applying JPQ index compression to the dense embedding index. It evaluates retrieval latency, index size, and downstream weighted F1 on HoVer, WiCE, and a newly released set of 281 claims from the 2024 US presidential debate. The main reported results are up to 10x CPU speedups, more than 20x GPU speedups, a 93% reduction in dense index size, and small F1 differences relative to a full-corpus sparse-retrieval-plus-rerank baseline.

Significance. If the speedups and F1 preservation hold, the paper addresses a practical bottleneck in fact-checking: many existing pipelines focus on verification while ignoring the cost of retrieving evidence from Wikipedia-scale corpora. The CPU speedups, the large index compression, and the open-sourced debate claims dataset are concrete, useful contributions, and the paper is refreshingly explicit about the retrieval-efficiency side of the pipeline. However, the GPU speedup headline is currently not supported by the reported baseline, and the paper leaves ambiguity about which evaluation split was used for the F1 results; these issues weaken the central quantitative claims and require correction before the results can be fully trusted.

major comments (4)
  1. [Table 3, Insight 3, Abstract] The abstract and Insight 3 claim 'more than a 20.0x speedup on GPUs,' but Table 3 does not provide a GPU total latency for the Full-Wiki (S+R) baseline: the GPU column for the two baseline rows is a dash. The reported GPU speedups (33.0x for HoVer, 18.1x for WiCE) are therefore computed as the CPU baseline total latency (659 ms and 831 ms) divided by the proposed GPU total latency (20 ms and 46 ms), which is a CPU-vs-GPU comparison. A same-device dense baseline from the same table (Full-Wiki Dense GPU totals of 174 ms and 238 ms) yields roughly 8.7x and 5.2x speedups for Citation Extraction, not 33x/18x. Please measure a GPU S+R (or same-device GPU) baseline, or state clearly that the GPU speedup is relative to a CPU baseline; the same issue affects Table 4's GPU speedups (e.g., 695/157 = 4.4x for Fusion).
  2. [Section 4.1 / Figure 2] HoVer and WiCE are described as having unlabeled test splits, yet Figure 2 reports weighted F1 scores. It should be stated explicitly whether these scores are computed on the development sets (presumably) or on a held-out portion with labels obtained by other means. Because the development-set label distribution and size differ from the test splits, the downstream 'minimal performance loss' claim cannot be fully assessed without this information.
  3. [Section 3.1 / Figure 3] The pruning methods remove 52–61% of sentences, so the speedup claims rest on the assumption that the retained facts contain essentially all evidence needed for verification. End-to-end F1 on HoVer and WiCE is evidence for this, but Figure 3 provides direct retrieval-recall evidence only for WiCE (nDCG@10, Recall@10); HoVer and the debate dataset lack a comparable retrieval-recall analysis. Please add sentence-level recall on HoVer and the debate data, and if possible a per-claim-type breakdown, to bound the risk that pruning discards evidence on particular claim distributions.
  4. [Tables 2–4, Figure 2] All latency and F1 numbers appear to be single measurements. The speedup ratios in Tables 2–4 are the central quantitative claims, and latency measurements are typically noisy; please report means and standard deviations over multiple runs (or at least multiple query batches) and indicate whether the F1 differences between Full-Wiki and the pruned/compressed setups are within measurement noise.
minor comments (6)
  1. [Title / running header] The paper is titled 'FlashCheck' but the running header and some text use 'FastCheck'; please unify the name throughout.
  2. [Figure 2] Figure 2 has no legend and the numeric labels are not clearly aligned with the bars; please redraw with named groups (e.g., Sparse+re-rank and Dense IC) and readable axis labels.
  3. [Section 4.2] The decision threshold for the ClaimBuster model used in Fact Extraction and Fusion is not reported; this threshold controls the amount of pruning and should be stated.
  4. [Section 5.3] The text states 'GPU retrieval shows about 2.0x and 0.8x speedups,' but Table 3 reports total-latency speedups; please clarify what retrieval-only GPU numbers are being referred to.
  5. [Section 5.4] The 2024 debate dataset would benefit from a description of how the 281 claims and their labels were produced, including annotator instructions and agreement.
  6. [Section 3.2] Please reconcile the compression-ratio formula 4D/M with the measured 93% reduction (14.4:1); the arithmetic is not immediate from the stated D and M (e.g., 4*384/96 = 16 for the MiniLM encoder), so clarify the exact byte-level calculation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the speedup and accuracy results are empirical measurements against external benchmarks, and self-citations appear only as related-work mentions.

full rationale

The paper's central claims are the speedups and F1 scores reported in Tables 2, 3, and 4 and in Insights 1-3. These are direct measurements of end-to-end pipeline latency and downstream verification performance on external datasets (HoVer, WiCE, and a newly annotated 2024 presidential debate set), not quantities derived from the paper's own assumptions by construction. The proposed pipeline combines a pruned factual-sentence index (via ClaimBuster-based claim detection and Wikipedia citation heuristics) with JPQ index compression, where JPQ is an external prior method cited as [34]. The acceleration and accuracy outcomes could have contradicted the pruning assumption, and the paper reports such outcomes transparently. The self-citations [17], [27], and [30] appear only in the related-work enumeration of automated fact-checking systems and efficient IR approaches; none of them is load-bearing for the paper's derivation or evaluation, and no uniqueness theorem or ansatz is imported from same-author prior work. The pruning step is a design assumption about which corpus content is useful for verification, but it is not circular: it is evaluated by downstream task F1 against external labels. The skeptic observation that Table 3's GPU speedups use a CPU classical baseline is a benchmarking/reporting concern about fair comparison, not a circularity pattern specified here, so it does not raise the circularity score. Overall, the derivation chain is self-contained and empirically grounded, with only minor non-load-bearing self-citations.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on corpus-pruning heuristics and JPQ compression hyperparameters. No new theoretical entities are proposed; the released debate dataset is a data artifact, not an invented construct. The free parameters are the JPQ codebook settings and the unreported claim-detection threshold. The axioms are domain assumptions about what makes Wikipedia content factual and about transferability of the claim-detection and JPQ models.

free parameters (3)
  • JPQ codebooks M = 96
    Number of subquantizers in product quantization; sets the 4D/M compression ratio and directly controls the speed-accuracy trade-off reported in Section 5.3.
  • JPQ codewords K = 256
    Number of centroids per codebook; chosen in Section 4.2 and affects index fidelity and memory.
  • ClaimBuster decision threshold = not reported
    The fact-extraction method (FE) depends on the claim-detection model's threshold for retaining sentences, but no threshold is stated in Section 4.2, leaving an unreported free parameter in the pruning pipeline.
assumptions (5)
  • domain assumption Sentences with Wikipedia citations are factual statements sufficient for verification
    The citation-extraction approach (CE) in Section 3.1 assumes that citation-backed sentences are the factual content needed for fact-checking and that uncited sentences can be pruned without losing evidence.
  • domain assumption ClaimBuster model identifies factual sentences relevant to claims
    The fact-extraction approach (FE) in Section 3.1 relies on a claim-detection model trained on ClaimBuster, assuming it transfers to Wikipedia text and debate claims.
  • domain assumption The pruned corpus preserves information needed by the downstream verifier
    The paper drops the sentence-selection stage because the pruned index supposedly contains succinct facts; the F1 results in Figure 2 test this indirectly but only on two datasets.
  • domain assumption JPQ jointly trained query encoder and quantized index preserves retrieval quality
    Section 3.2 borrows JPQ from prior work [34] and assumes the quantized index retains enough retrieval quality to keep downstream verification performance close to the uncompressed dense baseline.
  • domain assumption Wikipedia is a reliable knowledge source for fact-checking
    The entire retrieval pipeline uses only Wikipedia dumps as the knowledge source; claims or facts not present in Wikipedia cannot be verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlashCheck: Exploration of Efficient Evidence Retrieval for Fast Fact-Checking." pith.science (2026). https://pith.science/paper/QSDAENTI

@misc{pith2026250205803,
  author       = {Pith},
  title        = {Pith review of: FlashCheck: Exploration of Efficient Evidence Retrieval for Fast Fact-Checking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QSDAENTI}},
  note         = {Machine review of arXiv:2502.05803}
}
read the original abstract

The advances in digital tools have led to the rampant spread of misinformation. While fact-checking aims to combat this, manual fact-checking is cumbersome and not scalable. It is essential for automated fact-checking to be efficient for aiding in combating misinformation in real-time and at the source. Fact-checking pipelines primarily comprise a knowledge retrieval component which extracts relevant knowledge to fact-check a claim from large knowledge sources like Wikipedia and a verification component. The existing works primarily focus on the fact-verification part rather than evidence retrieval from large data collections, which often face scalability issues for practical applications such as live fact-checking. In this study, we address this gap by exploring various methods for indexing a succinct set of factual statements from large collections like Wikipedia to enhance the retrieval phase of the fact-checking pipeline. We also explore the impact of vector quantization to further improve the efficiency of pipelines that employ dense retrieval approaches for first-stage retrieval. We study the efficiency and effectiveness of the approaches on fact-checking datasets such as HoVer and WiCE, leveraging Wikipedia as the knowledge source. We also evaluate the real-world utility of the efficient retrieval approaches by fact-checking 2024 presidential debate and also open source the collection of claims with corresponding labels identified in the debate. Through a combination of indexed facts together with Dense retrieval and Index compression, we achieve up to a 10.0x speedup on CPUs and more than a 20.0x speedup on GPUs compared to the classical fact-checking pipelines over large collections.

Figures

Figures reproduced from arXiv: 2502.05803 by the authors.

Figure 1
Figure 1. Comparison of Existing and Proposed Fact-Checking Pipelines 3.1 Corpus Compression through Extraction of Facts Large collections like Wikipedia are usually employed as knowledge sources for fact-checking [12, 14]. While each Wikipedia article has detailed information on a topic, all the information is not factual, lacking citations and less informative for purposes of verifying claims. Hence, identifying useful fact… view at source ↗
Figure 2
Figure 2. HoVer and WiCe task performance (FW- Full-Wiki, FE - Fact Extraction, [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Retrieval performance comparison Fusion approaches, maintains performance closest to the Full-Wiki setup as measured by weighted F1 in [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Live fact-checking performance across different corpus setups [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 24 canonical work pages

  1. [1]

    Aly and A

    R. Aly and A. Vlachos. Natural logic-guided autoregressive multi-hop document retrieval for fact verification. In Y. Goldberg, Z. Kozareva, and Y. Zhang, editors, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6123– 6135, Abu Dhabi, United Arab Emirates, Dec. 2022. Association for Computational Linguis- tics....

    work page doi:10.18653/v1/2022.emnlp-main.411 2022

  2. [2]

    Baranchuk, A

    D. Baranchuk, A. Babenko, and Y. Malkov. Revisiting the inverted indices for billion-scale approximate nearest neighbors, 2018

    work page 2018

  3. [3]

    Bondarenko, M

    Y. Bondarenko, M. Nagel, and T. Blankevoort. Understanding and overcoming the challenges of efficient transformer quantization, 2021

    work page 2021

  4. [4]

    Bruch, F

    S. Bruch, F. M. Nardini, C. Rulli, and R. Venturini. Efficient inverted indexes for approximate retrieval over learned sparse representations. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’24, page 152–162, New York, NY, USA, 2024. Association for Computing Machinery. ISBN 979840...

    arXiv 2024

  5. [5]

    T. Ge, K. He, Q. Ke, and J. Sun. Optimized product quantization. IEEE Transactions on Pattern Analysis and Machine Intelligence , 36(4):744–755, 2014. https://doi.org/10.1109/TPAMI.2013.240

    doi:10.1109/tpami.2013.240 2014

  6. [6]

    J. Guo, Y. Cai, Y. Fan, F. Sun, R. Zhang, and X. Cheng. Semantic models for the first-stage retrieval: A comprehensive review. ACM Transactions on Information Systems, 40(4):1–42, Mar. 2022. ISSN 1558-2868. https://doi.org/10.1145/3486250. URL http://dx.doi.org/10.1145/ 3486250

    work page doi:10.1145/3486250 2022

  7. [7]

    Z. Guo, M. Schlichtkrull, and A. Vlachos. A survey on automated fact-checking. Transactions of the Association for Computational Linguistics, 10:178–206, 2022. URL https://aclanthology. org/2022.tacl-1.11

    work page 2022

  8. [8]

    Y. Han, C. Liu, and P . Wang. A comprehensive survey on vector database: Storage and retrieval technique, challenge, 2023

    work page 2023

Show all 39 references
  1. [9]

    Hassan, C

    N. Hassan, C. Li, and M. Tremayne. Detecting check-worthy factual claims in presidential debates. CIKM ’15. Association for Computing Machinery, 2015

    2015

  2. [10]

    Hofst¨ atter, S.-C

    S. Hofst¨ atter, S.-C. Lin, J.-H. Yang, J. Lin, and A. Hanbury . Efficiently teaching an effective dense retriever with balanced topic aware sampling. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’21,...

    2021

  3. [11]

    X. Hu, Z. Hong, Z. Guo, L. Wen, and P. Yu. Read it twice: Towards faithfully interpretable fact verification by revisiting evidence. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’23, page 2319–2323, N...

    2023

  4. [12]

    Jiang, S

    Y. Jiang, S. Bordia, Z. Zhong, C. Dognin, M. Singh, and M. Bansal. Hover: A dataset for many-hop fact extraction and claim verification, 2020

    2020

  5. [13]

    J´ egou, M

    H. J´ egou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1):117–128, 2011. https://doi.org/10.1109/TPAMI.2010.57

    2011 doi

  6. [14]

    Kamoi, T

    R. Kamoi, T. Goyal, J. Rodriguez, and G. Durrett. WiCE: Real-world entailment for claims in Wikipedia. In H. Bouamor, J. Pino, and K. Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7561–7583, Singapore, Dec. 2023. A...

    2023

  7. [15]

    Karpukhin, B

    V. Karpukhin, B. O˘ guz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W. tau Yih. Dense passage retrieval for open-domain question answering, 2020

    2020

  8. [16]

    Lazarski, M

    E. Lazarski, M. Al-Khassaweneh, and C. Howard. Using nlp for fact checking: A survey. Designs, 5(3), 2021. ISSN 2411-9660. https://doi.org/10.3390/designs5030042. URL https: //www.mdpi.com/2411-9660/5/3/42

    2021 doi

  9. [17]

    Leonhardt, H

    J. Leonhardt, H. M¨ uller, K. Rudra, M. Khosla, A. Anand, and A. Anand. Efficient neural ranking using forward indexes and lightweight encoders, 2023. URL https://arxiv.org/abs/2311.01263

    2023 arXiv

  10. [18]

    Li and Y

    Y. Li and Y. Xie. Is a picture worth a thousand words? an empirical study of image content and social media engagement. Journal of Marketing Research, 57, 2020

    2020

  11. [19]

    Y. Luan, J. Eisenstein, K. Toutanova, and M. Collins. Sparse, dense, and attentional represen- tations for text retrieval. Transactions of the Association for Computational Linguistics, 9: 329–345, 2021. URL https://aclanthology.org/2021.tacl-1.20

    2021

  12. [20]

    Nakov, D

    P. Nakov, D. Corney, M. Hasanain, F. Alam, T. Elsayed, A. Barr´ on-Cede˜ no, P. Papotti, S. Shaar, and G. D. S. Martino. Automated fact-checking for assisting human fact-checkers, 2021. URL https://arxiv.org/abs/2103.07769

    2021 arXiv

  13. [21]

    Newman, M

    E. Newman, M. Garry, D. Bernstein, J. Kantner, and D. Lindsay. Nonprobative photographs (or words) inflate truthiness. Psychonomic bulletin and review, 19, 2012. FastCheck: Fast fact-checking through efficient evidence retreival 15

    2012

  14. [22]

    L. Pan, X. Wu, X. Lu, A. T. Luu, W. Y. W ang, M.-Y. Kan, and P. Nakov. Fact-checking complex claims with program-guided reasoning. In A. Rogers, J. Boyd-Graber, and N. Okazaki, editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vo...

    2023 doi

  15. [23]

    Petroni, S

    F. Petroni, S. Broscheit, A. Piktus, P . Lewis, G. Izacard, L. Hosseini, J. Dwivedi-Y u, M. Lomeli, T. Schick, P.-E. Mazar´ e, A. Joulin, E. Grave, and S. Riedel. Improving wikipedia verifiability with ai, 2022

    2022

  16. [24]

    Samarinas, W

    C. Samarinas, W. Hsu, and M. L. Lee. Improving evidence retrieval for automated explainable fact-checking. In A. Sil and X. V. Lin, editors, Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Techno...

    2021 doi

  17. [25]

    M. S. Schlichtkrull, V. Karpukhin, B. Oguz, M. Lewis, W.-t. Yih, and S. Riedel. Joint verification and reranking for open fact checking over tables. In C. Zong, F. Xia, W. Li, and R. Navigli, editors, Proceedings of the 59th Annual Meeting of the Association for Computational ...

    2021 doi

  18. [26]

    Schuster, A

    T. Schuster, A. Fisch, and R. Barzilay. Get your vitamin c! robust fact verification with contrastive evidence, 2021. URL https://arxiv.org/abs/2103.08541

    2021 arXiv

  19. [27]

    Setty and V

    R. Setty and V. Setty. Questgen: Effectiveness of question generation methods for fact-checking applications. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM ’24, page 4036–4040, New York, NY, USA, 2024. Association for Com...

    2024

  20. [28]

    Thorne and A

    J. Thorne and A. Vlachos. Automated fact checking: Task formulations, methods and future directions, 2018

    2018

  21. [29]

    Thorne, A

    J. Thorne, A. Vlachos, C. Christodoulopoulos, and A. Mittal. Fever: a large-scale dataset for fact extraction and verification, 2018

    2018

  22. [30]

    Venktesh, A

    V. Venktesh, A. Anand, A. Anand, and V. Setty. Quantemp: A real-world open-domain benchmark for fact-checking numerical claims, 2024. URL https://arxiv.org/abs/2403.17169

    2024 arXiv

  23. [31]

    Z. Wei, X. Xu, C. Wang, Z. Liu, P. Xin, and W. Zhang. An index construc- tion and similarity retrieval method based on sentence-bert. In 2022 7th Interna- tional Conference on Image, Vision and Computing (ICIVC) , pages 934–938, 2022. https://doi.org/10.1109/ICIVC55077.2022.9886134

    2022

  24. [32]

    Yin and D

    W. Yin and D. Roth. TwoWingOS: A two-wing optimization strategy for evidential claim verification. In E. Riloff, D. Chiang, J. Hockenmaier, and J. Tsujii, editors, Pro- ceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 105–114, Brussels...

    2018 doi

  25. [33]

    J. Zhan, J. Mao, Y. Liu, M. Zhang, and S. Ma. Repbert: Contextualized text embeddings for first-stage retrieval, 2020. URL https://arxiv.org/abs/2006.15498

    2020 arXiv

  26. [34]

    J. Zhan, J. Mao, Y. Liu, J. Guo, M. Zhang, and S. Ma. Jointly optimizing query encoder and product quantization to improve retrieval performance, 2021

    2021

  27. [35]

    W. X. Zhao, J. Liu, R. Ren, and J.-R. Wen. Dense text retrieval based on pretrained language models: A survey, 2022

    2022

  28. [36]

    X. Zhao, Y. Tian, K. Huang, B. Zheng, and X. Zhou. Towards efficient index construction and approximate nearest neighbor search in high-dimensional spaces. Proc. VLDB Endow., 16(8):1979–1991, jun 2023. ISSN 2150-8097. https://doi.org/10.14778/3594512.3594527. URL https://doi.o...

    1979

  29. [37]

    Zheng, C

    L. Zheng, C. Li, X. Zhang, Y.-M. Shang, F. Huang, and H. Jia. Evidence retrieval is almost all you need for fact verification. In L.-W. Ku, A. Martins, and V. Srikumar, editors, Findings of the As- sociation for Computational Linguistics: ACL 2024, pages 9274–9281, Bangkok, Th...

    2024

  30. [38]

    X. Zhu, J. Li, Y. Liu, C. Ma, and W. Wang. A survey on model compression for large language models, 2023

    2023

  31. [2024]

    https://doi.org/10.18653/v1/2024.findings- acl.551

    Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.findings- acl.551. URL https://aclanthology.org/2024.findings-acl.551

    2024 doi

Pith tools

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