Pith. sign in

REVIEW 3 major objections 6 minor 41 references

Multimodal rumor detection enhanced by external evidence and forgery features

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read A multimodal rumor detector that adds Fourier-based forgery features and BLIP captions to external evidence claims to beat all compared baselines on MR2.

desk verdict A reasonable engineering combination with solid ablations, but the headline SOTA rests on an unfair baseline comparison—DEETSA is missing and the reported 'Macro-F1' is actually Macro Accuracy. read the letter →

arxiv 2601.14954 v3 pith:PQKJHNN2 submitted 2026-01-21 cs.LG

classification cs.LG
keywords multimodalrumordetectionimage-textsemanticmatchingmodalityfusionattentionmechanismimageforgeryexternalevidencecontrastivelearningBLIPcaptioning
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

Social-media rumors increasingly pair text with images that are either subtly forged or only superficially consistent, and existing multimodal detectors often miss such cases because they ignore external verification and tampering clues. This paper aims to fix that by adding two signals to an evidence-augmented model: frequency-domain traces of image manipulation and BLIP-generated captions that serve as a concise, news-like semantic bridge between image and text. The model tightens alignment with two contrastive losses, one between text and image and one between text and image description, and fuses everything through gated adaptive scaling. On the MR2 benchmark it reports Macro-F1 of 94.9% on Weibo and 94.1% on Twitter, beating all six compared baselines; ablations point to the gating module and the forgery module as the biggest contributors.

What carries the argument

The load-bearing objects: (1) a forgery-feature module running 2D discrete Fourier transform, amplitude spectrum, low-frequency truncation, multi-scale convolution, and multi-head self-attention to catch tampering traces; (2) BLIP captioning, which turns images into short, news-like descriptions, creating a linguistic bridge; (3) dual InfoNCE losses aligning text–image and text–description pairs; and (4) hierarchical gated fusion with adaptive scaling, blending original post, external evidence, and forgery features.

What would settle it

Run the six baselines and the cited dual-evidence model under identical conditions — same evidence, same data split, same hyperparameter search budget — and check whether the Macro-F1 gap shrinks below the reported ~1-2 point advantage. A second check: remove the forgery module but retrain everything else; if the drop on Twitter is less than the reported 3.0% macro-accuracy loss, the forgery feature's causal role is unsupported.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims a multimodal rumor detector that integrates four mechanisms: a forgery-feature module applying a 2D Fourier transform, low-frequency truncation, multi-scale convolutions, and multi-head self-attention; BLIP captions that bridge visual and textual semantics with concise descriptions; dual InfoNCE losses aligning text–image and text–description pairs; and hierarchical gated fusion with adaptive feature scaling, all fed by external text and image evidence via multi-head attention. Reported results give 94.9% Macro-F1 on Weibo and 94.1% on Twitter, above the six compared baselines, with ablations crediting the gating mechanism and forgery features with the larg

Load-bearing premise

The comparison is fair: the paper does not state whether the non-evidence baselines were given the same external evidence or tuned under identical conditions, and the closest prior work that already uses dual evidence enhancement and BLIP captions, cited in the related-work section, is absent from the comparison tables.

Editorial extensions

If this is right

  • If the reported numbers are accurate, incorporating frequency-domain forgery analysis into multimodal rumor detectors provides signal not captured by semantic-only or event-level models.
  • The optimal evidence count of 5 on both datasets implies that evidence-augmented detection should treat retrieval quantity as a tunable hyperparameter rather than assuming more evidence is always better.
  • BLIP-style captions, being concise and close to news style, may reduce semantic noise in cross-modal alignment compared to open-ended caption generators.
  • The reported cross-lingual consistency (94.9% on Chinese Weibo vs 94.1% on English Twitter) suggests the approach is not tied to a specific language or platform.

Reading between the lines

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

  • If the architecture's advantage truly comes from forgery features, a natural testable extension is to apply the same Fourier-based module to other multimodal classification tasks, such as detecting AI-generated images or manipulated product photos, where tampering traces matter.
  • The evidence-quantity peak at 5 may indicate that an adaptive evidence-ranking or de-noising mechanism could push performance further, since the paper currently uses a fixed maximum per post.
  • The reliance on BLIP for captions suggests sensitivity to captioner quality; a cheap experiment would replace BLIP with a smaller captioner or a zero-shot vision-language model and measure whether the dual-contrast gain persists.
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

3 major / 6 minor

Summary. The paper proposes a multimodal rumor detection model for the MR2 benchmark. The architecture combines a ResNet34 image encoder, a BERT text encoder, a Fourier-based forgery feature module, BLIP-generated image descriptions with dual text-image and text-description contrastive losses, evidence attention over retrieved textual and visual evidence, and a gated adaptive fusion mechanism. The authors report macro accuracy/F1 of 94.9% on Weibo and 94.1% on Twitter, outperforming six baselines (CLIP, LSTM_word2vec, DEDA, HMCAN, CCN, MRAN). Ablations over seven variants and an evidence-quantity sweep are also reported. The central claim is empirical: that this architecture achieves state-of-the-art performance on MR2.

Significance. If the empirical claim is sound, the contribution is a competitive integrated system for multimodal rumor detection, combining several previously separate ideas (forgery features, BLIP-based contrastive alignment, external evidence attention, gated fusion) in one framework. The use of the external MR2 benchmark, the seven-variant ablation study, and the evidence-quantity sensitivity analysis are strengths. The main weakness is that the state-of-the-art claim rests on an incomplete comparison protocol: the closest prior method, DEETSA [27], is cited as inspiration but excluded from all comparative tables, and the paper does not document whether baselines were given identical evidence, preprocessing, and tuning. The reported headline metric is also inconsistently defined. The architecture itself is plausible and the internal ablations are generally consistent, but the comparative evidence is currently insufficient to support the SOTA claim.

major comments (3)
  1. [§4.3.2, Tables 1–2] The central claim that the proposed method outperforms mainstream baselines depends on fair comparison, but the paper does not state whether CLIP, LSTM_word2vec, DEDA, HMCAN, CCN, or MRAN were given the same external evidence, the same 8:1:1 split, or the same hyperparameter tuning. Several of these methods predate MR2 and were likely re-implemented by the authors, yet no code or configuration is released. More importantly, DEETSA [27], which is explicitly credited in §3.2 and §3.3.1 as the inspiration for the evidence-attention and BLIP-based alignment modules, is absent from Tables 1–2 and from all ablations. DEETSA already combines dual evidence enhancement, BLIP captions, text–image similarity awareness, and adaptive gated fusion. Without a direct comparison to DEETSA, or at least evidence that all baselines were run under identical, adequately tuned conditions, the reported 94.9/94.
  2. [§4.3.2, Tables 1–2] The metric used for the headline numbers is inconsistently described. The text states that the method 'reaches a Macro-F1 of 94.9% on the Weibo dataset and 94.1% on the Twitter dataset,' but the corresponding columns in Tables 1 and 2 are labeled 'Macro Accuracy.' The abstract also claims improvement in 'macro accuracy, recall, and F1 score,' but no single macro-recall or macro-F1 is defined or reported. Recomputing macro-F1 from the per-class F1 columns gives approximately 94.2 (Weibo) and 93.7 (Twitter), which are not the stated values. The authors should clarify exactly which macro metric is used and ensure the text, tables, and abstract are mutually consistent.
  3. [§4.3.3, Tables 3–4 (labeled 'Table 1' and 'Table 2')] The ablation study supports the contribution of individual modules, but the reported differences are small (typically 0.2–3.5 macro-accuracy points), and no variance, repeated runs, or statistical significance tests are provided. Since the model and baselines appear to be trained a single time with early stopping, the observed differences between w/o T-T DC and the full model (0.8 on Weibo, 0.9 on Twitter) could lie within run-to-run noise. The same concern applies to the evidence-quantity sweep in Figure 2, which plots single runs. To make the comparative and ablation claims robust, the authors should report mean and standard deviation over multiple seeds, or at least provide error bars and a significance test.
minor comments (6)
  1. [§4.3.2] The text contains unresolved reference placeholders '错误!未找到引用源。' instead of table citations. Please replace them with proper cross-references to Tables 1 and 2.
  2. [§4.3.3] The ablation tables are captioned 'Table 1' and 'Table 2', duplicating the captions of the main comparison tables. They should be labeled Table 3 and Table 4.
  3. [§4.1] The text says 'The WR2 datasets are introduced first,' but the dataset is MR2. Please fix the typo.
  4. [§3.4.2] The notation ℎ_T' and ℎ_I' is reused: previously in §3.2 these symbols denote evidence-attention outputs, and in §3.4.2 they are redefined as linearly projected fusion features. Please disambiguate the notation to avoid confusion.
  5. [§3.5] The cross-entropy loss formula is written with p_i as the ground-truth probability and q_i as the predicted probability, but the roles are reversed in the standard notation (p is predicted, y is target). Please align the notation with the text.
  6. [§3.4.3] The adaptive scaling parameter λ is defined as learnable and bounded in (0,1], but no mechanism is described for enforcing that bound. If a sigmoid or similar constraint is used, please state it explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical benchmark evaluation with external encoders and standard losses; reported results are held-out test metrics.

full rationale

This paper is an empirical systems paper rather than a derivation, and no load-bearing step reduces to its own input. Text/image/forgery features are computed with external models (BERT, ResNet34, BLIP, Fourier transforms); the dual-contrastive losses, gated fusion, and adaptive scaling (α, β, γ, λ) are trainable components optimized on labeled MR2 training data; the headline Weibo/Twitter numbers are test-set evaluations, not consequences of the model definition. There are no self-citations: reference [27] (DEETSA, Huang et al.) and [28] (BLIP) are independent external works, and no uniqueness theorem or ansatz is imported from the authors' own prior output. The learnable scalars are not pre-fitted to test labels, and no fitted parameter is renamed as a prediction. The omission of DEETSA from Tables 1–2, the unspecified baseline tuning/evidence conditions, and the Macro-F1 versus Macro Accuracy labeling inconsistency are legitimate experimental-fairness and reporting concerns, but they are not circularity: they concern whether the comparison is convincing, not whether the model's output is contained in its inputs.

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

The central claim rests on pretrained encoders, BLIP captions, MR2 evidence, and an equitable baseline comparison; no new physical or ontological entities are introduced, but several important hyperparameters are unreported and the baseline fairness assumption is unverified.

free parameters (5)
  • InfoNCE temperature τ
    Appears in L_TT and L_TI (§3.3.2, §3.3.3); no value is reported, and it controls how hard negative pairs are weighted.
  • Contrastive balancing coefficients λ_TT, λ_TI
    Weights in the joint loss L_sim = λ_TT L_TT + λ_TI L_TI (§3.3.4); values are not reported.
  • Adaptive feature-scaling scalar λ
    Learned scalar in (0,1] multiplying the fused features (§3.4.3); value is data-dependent and not reported.
  • Low-frequency truncation radius R_low
    Defines the circular low-frequency region Ω_low in the forgery module (§3.1.3); hand-chosen and not reported.
  • Maximum evidence items per post = 5
    Selected from a sweep of 1 to 9 evidence items on Weibo and Twitter (§4.3.1, Figure 2), then used in the final model; data-dependent hyperparameter.
assumptions (5)
  • domain assumption BLIP descriptions are faithful image summaries and stylistically close to news text
    Used as the semantic bridge in §3.3.1; if captions misdescribe images, the text-description contrastive loss encodes wrong alignment.
  • domain assumption MR2's retrieved textual/visual evidence is relevant and does not leak labels
    The evidence fusion module (§3.2) depends on the public MR2 evidence; unreliable or label-leaking evidence would inflate results independently of the architecture.
  • domain assumption Low-frequency removal preserves useful manipulation traces in the amplitude spectrum
    The forgery module (§3.1.3) assumes tampering artifacts survive after zeroing low-frequency components; no independent forgery-dataset validation is provided.
  • domain assumption Cross-modal inconsistency is a usable rumor signal even when rumor makers select matched images
    The dual-contrast module (§3.3) relies on text-image/text-description mismatch, while the paper itself cites Zhang et al. [24] arguing that rumors often use highly matched images.
  • standard math Standard DFT, softmax attention, InfoNCE, and pretrained encoders behave as specified
    Background mathematical machinery in §3.1.3, §3.2, and §3.3; assumed without proof, which is normal for applied ML.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal rumor detection enhanced by external evidence and forgery features." pith.science (2026). https://pith.science/paper/PQKJHNN2

@misc{pith2026260114954,
  author       = {Pith},
  title        = {Pith review of: Multimodal rumor detection enhanced by external evidence and forgery features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PQKJHNN2}},
  note         = {Machine review of arXiv:2601.14954}
}
read the original abstract

Social media increasingly disseminates information through mixed image text posts, but rumors often exploit subtle inconsistencies and forged content, making detection based solely on post content difficult. Deep semantic mismatch rumors, which superficially align images and texts, pose particular challenges and threaten online public opinion. Existing multimodal rumor detection methods improve cross modal modeling but suffer from limited feature extraction, noisy alignment, and inflexible fusion strategies, while ignoring external factual evidence necessary for verifying complex rumors. To address these limitations, we propose a multimodal rumor detection model enhanced with external evidence and forgery features. The model uses a ResNet34 visual encoder, a BERT text encoder, and a forgery feature module extracting frequency domain traces and compression artifacts via Fourier transformation. While some existing approaches employ large scale generative vision language models for caption generation, their open ended generation tendency produces verbose and stylistically inconsistent descriptions that introduce semantic noise and risk drifting from actual image content. To overcome this, we adopt BLIP specifically pre trained for vision language alignment which generates concise, image faithful descriptions stylistically closer to news text, serving as a reliable semantic bridge across modalities. A BLIP Driven Semantic Alignment Module jointly optimizes text image and text description contrastive losses, capturing inconsistencies at both visual and semantic levels. A gated adaptive feature scaling fusion mechanism dynamically adjusts multimodal fusion and reduces redundancy. Experiments on Weibo and Twitter datasets demonstrate that our model outperforms mainstream baselines in,recall,and F1 score

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 6 canonical work pages

  1. [27]

    Multimodal detection of information disorder from social media

    Kirchknopf A, Slijepčević D, Zeppelzauer M. Multimodal detection of information disorder from social media. In: Proceedings of the International Conference on Content -Based Multimedia Indexing; 2021 Jun. p. 1–4. https://doi.org/10.1109/CBMI50390.2021.9485012

  2. [1]

    w/o D C: The Dual contrastive learning module is removed

  3. [2]

    w/o Forgery: The forgery feature extraction module is removed

  4. [3]

    w/o Gating: The gated fusion and adaptive feature scaling mechanisms are removed

  5. [4]

    The WR2 datasets are introduced first, followed by implementation settings and evaluation metrics

    Experiments This section outlines the experimental protocol used to assess the proposed method. The WR2 datasets are introduced first, followed by implementation settings and evaluation metrics. Results are then reported for hyperparameter sensitivity, state -of-the-art comparisons, ablations, and error analysis. A final case study illustrates practical s...

  6. [5]

    w/o Evidence: Both textual and visual evidence are removed

  7. [6]

    w/o T-I D C: The text–image contrastive learning module is removed

  8. [7]

    w/o T-T D C: The text–image description contrast module is removed

Show all 41 references
  1. [8]

    The results on the Weibo and Twitter datasets are shown in Tables 3 and 4, respectively

    w/o F S: Only the adaptive feature scaling mechanism is removed. The results on the Weibo and Twitter datasets are shown in Tables 3 and 4, respectively. By comparing the full model with these seven variants, the ablation study effectively validates the necessity and effective...

  2. [9]

    Experiments on the MR2 dataset show improvements in multimodal reasoning, image forgery sensitivity, and cross-modal semantic alignment

    Conclusion This paper proposes a rumor detection method based on the enhancement of external evidence and forgery features, and significantly improves the model's performance in rumor detection tasks by introducing a dual contrastive learning mechanism and a gated modulation m...

  3. [10]

    A comprehensive survey of multimodal fake news detection techniques: Advances, challenges, and opportunities

    Tufchi S, Yadav A, Ahmed T. A comprehensive survey of multimodal fake news detection techniques: Advances, challenges, and opportunities. Int J Multimed Inf Retr. 2023;12(2):28. https://doi.org/10.1007/s13735-023-00296-3

  4. [11]

    Hierarchical fine -grained image forgery detection and localization

    Guo X, Liu X, Ren Z, Grosz S, Masi I, Liu X. Hierarchical fine -grained image forgery detection and localization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2023. p. 3155–3165. https://doi.org/10.1109/CVPR52729.2023.00308

  5. [12]

    Frequency domain filtered residual network for deepfake detection

    Wang B, Wu X, Tang Y, Ma Y, Shan Z, Wei F. Frequency domain filtered residual network for deepfake detection. Mathematics. 2023;11(4):816. https://doi.org/10.3390/math11040816

  6. [13]

    Cross -modal contrastive learning for multimodal fake news detection

    Wang L, Zhang C, Xu H, Xu Y, Xu X, Wang S. Cross -modal contrastive learning for multimodal fake news detection. In: Proceedings of the 31st ACM International Conference on Multimedia; 2023 Oct. p. 5696–5704. https://doi.org/10.1145/3581783.3613850

  7. [14]

    Cross -modal consistency learning with fine -grained fusion network for multimodal fake news detection

    Li J, Bin Y, Zou J, Wei J, Wang G, Yang Y. Cross -modal consistency learning with fine -grained fusion network for multimodal fake news detection. In: Proceedings of the 5th ACM International Conference on Multimedia in Asia; 2023 Dec. p. 1–7. https://doi.org/10.1145/3595916.3626397

  8. [15]

    Prominent features of rumor propagation in online social media

    Kwon S, Cha M, Jung K, Chen W, Wang Y. Prominent features of rumor propagation in online social media. In: Proceedings of the 2013 IEEE 13th International Conference on Data Mining; 2013 Dec. p. 1103–1108. https://doi.org/10.1109/ICDM.2013.61

  9. [16]

    Detecting rumors from microblogs with recurrent neural networks

    Ma J, Gao W, Mitra P, Kwon S, Jansen BJ, Wong KF, et al. Detecting rumors from microblogs with recurrent neural networks. In: Proceedings of the 25th International Joint Conference on Artificial Intelligence; 2016. p. 3818–3824

  10. [17]

    Call attention to rumors: Deep attention based recurrent neural networks for early rumor detection

    Chen T, Li X, Yin H, Zhang J. Call attention to rumors: Deep attention based recurrent neural networks for early rumor detection. In: Pacific-Asia Conference on Knowledge Discovery and Data Mining; 2018. p. 40–52. https://doi.org/10.1007/978-3-030-04503-6_4

  11. [18]

    Detecting breaking news rumors of emerging topics in social media

    Alkhodair SA, Ding SH, Fung BC, Liu J. Detecting breaking news rumors of emerging topics in social media. Inf Process Manag. 2020;57(2):102018. https://doi.org/10.1016/j.ipm.2019.02.016

  12. [19]

    Dual emotion based fake news detection: A deep attention- weight update approach

    Luvembe AM, Li W, Li S, Liu F, Xu G. Dual emotion based fake news detection: A deep attention- weight update approach. Inf Process Manag. 2023;60(4):103354. https://doi.org/10.1016/j.ipm.2023.103354

  13. [20]

    Near real -time topic -driven rumor detection in source microblogs

    Xu F, Sheng VS, Wang M. Near real -time topic -driven rumor detection in source microblogs. Knowl Based Syst. 2020;207:106391. https://doi.org/10.1016/j.knosys.2020.106391

  14. [21]

    Exploring the role of visual content in fake news detection

    Cao J, Qi P, Sheng Q, Yang T, Guo J, Li J. Exploring the role of visual content in fake news detection. In: Disinformation, Misinformation, and Fake News in Social Media: Emerging Research Challenges and Opportunities. 2020. p. 141–161. https://doi.org/10.1007/978-3-030-42699-6_8

  15. [22]

    Exploiting multi -domain visual information for fake news detection

    Qi P, Cao J, Yang T, Guo J, Li J. Exploiting multi -domain visual information for fake news detection. In: Proceedings of the 2019 IEEE International Conference on Data Mining; 2019 Nov. p. 518–527. https://doi.org/10.1109/ICDM.2019.00062

  16. [23]

    FakeGPT: Fake news generation, explanation and detection of large language models

    Huang Y, Sun L. FakeGPT: Fake news generation, explanation and detection of large language models. arXiv. 2023;arXiv:2310.05046. https://doi.org/10.48550/arXiv.2310.05046

  17. [24]

    Towards LLM -based fact verification on news claims with a hierarchical step - by-step prompting method

    Zhang X, Gao W. Towards LLM -based fact verification on news claims with a hierarchical step - by-step prompting method. In: Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia -Pacific Chapter of the Associa...

  18. [25]

    EANN: Event adversarial neural networks for multi-modal fake news detection

    Wang Y, Ma F, Jin Z, Yuan Y, Xun G, Jha K, et al. EANN: Event adversarial neural networks for multi-modal fake news detection. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining; 2018 Jul. p. 849 –857. https://doi.org/10.1145/3...

  19. [26]

    Fake news detection using deep Markov random fields

    Nguyen DM, Do TH, Calderbank R, Deligiannis N. Fake news detection using deep Markov random fields. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics; 2019 Jun. p. 1391 –1400. https://doi.org/10.18653/v1/n19-1142

  20. [28]

    CAF -ODNN: Complementary attention fusion with optimized deep neural network for multimodal fake news detection

    Luvembe AM, Li W, Li S, Liu F, Wu X. CAF -ODNN: Complementary attention fusion with optimized deep neural network for multimodal fake news detection. Inf Process Manag. 2024;61(3):103653. https://doi.org/10.1016/j.ipm.2024.103653

  21. [29]

    MVACLNet: A multimodal virtual augmentation contrastive learning network for rumor detection

    Liu X, Pang M, Li Q, Zhou J, Wang H, Yang D. MVACLNet: A multimodal virtual augmentation contrastive learning network for rumor detection. Algorithms. 2024;17(5):199. https://doi.org/10.3390/a17050199

  22. [30]

    Similarity -aware multi -modal fake news detection

    Zhou X, Wu J, Zafarani R. Similarity -aware multi -modal fake news detection. In: Pacific -Asia Conference on Knowledge Discovery and Data Mining; 2020. p. 354 –367. https://doi.org/10.1007/978-3-030-47436-2_27

  23. [31]

    Detecting fake news by exploring the consistency of multimodal data

    Xue J, Wang Y, Tian Y, Li Y, Shi L, Wei L. Detecting fake news by exploring the consistency of multimodal data. Inf Process Manag. 2021;58(5):102610. https://doi.org/10.1016/j.ipm.2021.102610

  24. [32]

    Open-domain, content-based, multi-modal fact-checking of out-of- context images via online resources

    Abdelnabi S, Hasan R, Fritz M. Open-domain, content-based, multi-modal fact-checking of out-of- context images via online resources. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022. p. 14940–14949

  25. [33]

    Multimodal fake news analysis based on image –text similarity

    Zhang X, Dadkhah S, Weismann AG, Kanaani MA, Ghorbani AA. Multimodal fake news analysis based on image –text similarity. IEEE Trans Comput Soc Syst. 2023;11(1):959 –972. https://doi.org/10.1109/TCSS.2023.3241571

  26. [34]

    Evidence-aware multimodal Chinese social media rumor detection

    Wu K, Cao D. Evidence-aware multimodal Chinese social media rumor detection. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing; 2024 Apr. p. 8376–8380. https://doi.org/10.1109/ICASSP48485.2024.10447346

  27. [35]

    Dual -channel early rumor detection based on factual evidence

    Wu Y, Sun J, Yuan X, Huang Z, Dai J. Dual -channel early rumor detection based on factual evidence. Expert Syst Appl. 2024;238:121928. https://doi.org/10.1016/j.eswa.2023.121928"

  28. [36]

    Dual evidence enhancement and text –image similarity awareness for multimodal rumor detection

    Huang X, Ma T, Rong H, Jia L, Su Y. Dual evidence enhancement and text –image similarity awareness for multimodal rumor detection. Eng Appl Artif Intell. 2025;153:110845. https://doi.org/10.1016/j.engappai.2025.110845

  29. [37]

    BLIP: Bootstrapping language -image pre-training for unified vision - language understanding and generation

    Li J, Li D, Xiong C, Hoi S. BLIP: Bootstrapping language -image pre-training for unified vision - language understanding and generation. In: Proceedings of the International Conference on Machine Learning; 2022 Jun. p. 12888–12900

  30. [38]

    MR2: A benchmark for multimodal retrieval -augmented rumor detection in social media

    Hu X, Guo Z, Chen J, Wen L, Yu PS. MR2: A benchmark for multimodal retrieval -augmented rumor detection in social media. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval; 2023 Jul. p. 2901 –2912. https://doi.or...

  31. [39]

    Learning transferable visual models from natural language supervision

    Radford A, Kim JW, Hallacy C, Ramesh A, Goh G, Agarwal S, et al. Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning; 2021 Jul. p. 8748–8763

  32. [40]

    Hierarchical multi -modal contextual attention network for fake news detection

    Qian S, Wang J, Hu J, Fang Q, Xu C. Hierarchical multi -modal contextual attention network for fake news detection. In: Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval; 2021. p. 153 –162. https://doi.org/10.1145/3...

  33. [41]

    MRAN: Multimodal relationship -aware attention network for fake news detection

    Yang H, Zhang J, Zhang L, Cheng X, Hu Z. MRAN: Multimodal relationship -aware attention network for fake news detection. Comput Stand Interfaces. 2024;89:103822. https://doi.org/10.1016/j.csi.2023.103822

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.