Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

KEN: Knowledge Augmentation and Emotion Guidance Network for Multimodal Fake News Detection

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read KEN, a multimodal fake news detector, augments images with LVLM captions and text with web-retrieved evidence, then routes news through emotion-specific detectors; it reports accuracy 0.935 on Weibo and 0.934 on Twitter, beating all ten…

desk verdict Solid knowledge-augmentation system paper whose emotion-guidance claim is undercut by a circular auxiliary loss that trains on the veracity label itself. read the letter →

arxiv 2507.09647 v2 pith:FPHVXILM submitted 2025-07-13 cs.MM cs.AI

Peican Zhu , Yubo Jing , Le Cheng , Keke Tang , Yangming Guo This is my paper · ORCID
classification cs.MMcs.AI
keywords multimodalfakenewsdetectionknowledgeaugmentationLVLMcaptioningevidenceretrievalemotion-guidedbalancedlearningCLIPfusionco-attentionWeiboandTwitterdatasets
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 proposes KEN, a multimodal fake news detector that tries to fix two gaps: models do not understand image semantics or obtain text context beyond the post, and they treat all news alike regardless of emotional tone. KEN therefore augments the image with a caption produced by a large vision-language model and augments the text with web-retrieved evidence, then fuses these through co-attention and CLIP-aligned embeddings. It also extracts emotion features from text and image, routes news through specialized detectors for five emotion types via a gating mechanism, and adds an auxiliary loss that assumes negative news is fake and positive news is real. On Weibo and Twitter, the paper reports accuracy around 0.935 and 0.934, with ablations indicating each added component contributes.

What carries the argument

The load-bearing mechanism is the pair of Knowledge Augmentation and Emotion Guidance. Knowledge Augmentation uses two co-attention blocks, one between text and web-retrieved evidence and one between image and generated caption, plus CLIP-enhanced embeddings $t_c$ and $v_c$ that map both modalities into a shared semantic space; the final fusion feature is $M_f = \theta \cdot \sigma_f(\mathrm{concat}(S_1, S_2, S_3))$, where $\theta = \cos(t_c, v_c)$ down-weights weak cross-modal matches. Emotion Guidance builds emotion features $M_e = \sigma_e(\mathrm{concat}(\gamma e_t, (1-\gamma) e_v))$ from multi-expert Bi-LSTM and self-attention analysis, then routes the aggregated news features through five emotion-domain processors weighted by a gating network. The auxiliary emotion reasoning loss $L_{\mathrm{emo}}$ uses the authenticity label $y$ as the emotion label under the assumption that negative means fake and positive means real, and the total loss is $L = L_{\mathrm{fnd}} + \lambda L_{\mathrm{emo}}$.

What would settle it

Annotate the emotional polarity of a sample of Weibo and Twitter posts independently of veracity, then measure the agreement between polarity and the authenticity label. If the agreement is near chance, Eq. (7)'s $L_{\mathrm{emo}}$ is not an emotion signal, and KEN's emotion-guidance ablations should be re-run with the emotion term replaced by a label-free emotion classifier; if accuracy does not drop, the guidance claim is unsupported.

Watch

Extended reading notes

Core claim

The central claim is that multimodal fake news detection improves when the model is given external knowledge and emotion-conditioned processing. For the image, a BLIP-2 caption supplies scene-level semantics that pixel-level features miss; for the text, GPT-4 with web search supplies evidence that resolves ambiguous terms and breaks the information silo of the post alone. Three fusion perspectives (text-evidence, image-caption, and CLIP-aligned image-text features) are combined with a similarity-weighted fusion that preserves unimodal features when cross-modal correlation is weak. Separately, emotion features from text and image are used to gate five emotion-specific processors, and an emotion reasoning loss is added to the main veracity loss. The paper reports KEN reaching accuracy 0.935 on Weibo and 0.934 on Twitter, above all ten compared baselines, with ablation studies attributing the gains to knowledge augmentation, CLIP alignment, balanced learning, gating, and emotion reasoning.

Load-bearing premise

The load-bearing premise is that a post's authenticity label can stand in for its emotional polarity: the emotion reasoning loss treats 'fake' as negative and 'real' as positive. If that correlation is weak, the auxiliary task is just a second veracity classifier, and the reported gains cannot be attributed to emotion modeling.

Editorial extensions

If this is right

  • If KEN's reported numbers hold, adding caption and evidence augmentation should improve any multimodal detector that currently relies only on raw pixels and post text.
  • Emotion-conditioned routing rather than one shared classifier appears to give a consistent accuracy gain on both datasets, with ablations showing 0.917 versus 0.935 on Weibo and 0.912 versus 0.934 on Twitter when emotion guidance is removed.
  • CLIP alignment contributes more than the whole knowledge augmentation module on Twitter (accuracy drops to 0.889 when CLIP is removed), suggesting cross-modal semantic alignment is a large part of the gain.
  • The optimal settings of five emotion types and three emotion-analysis experts on both datasets indicate a sweet spot; performance degrades when emotion categories are too coarse or too fine.

Reading between the lines

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

  • (Editorial inference) The emotion reasoning loss may actually be a second veracity classifier: Eq. (7) uses the authenticity label $y$ as the emotion label, so the reported w/o ER ablation may be measuring multi-task regularization rather than genuine emotion modeling.
  • (Editorial inference) KEN's evidence retrieval depends on GPT-4 and live web search, so results could be sensitive to retrieval freshness; on fast-moving events, the knowledge augmentation component may behave differently than on static benchmark datasets.
  • (Editorial inference) The balanced-learning idea could transfer to other classification tasks with known label-emotion correlations, such as rumor stance detection or review credibility, where emotion domains act as a soft mixture-of-experts routing.
  • (Editorial inference) A testable extension is to replace the veracity-derived emotion labels in $L_{\mathrm{emo}}$ with human or LLM emotion annotations; if accuracy is preserved under true emotion labels, the emotion-guidance claim would be substantially strengthened.
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 / 7 minor

Summary. The paper proposes KEN, a multimodal fake news detection network that augments news content with LVLM-generated image captions and web-retrieved textual evidence, fuses features through co-attention and CLIP-enhanced semantics, and incorporates an emotion analysis module with balanced learning over multiple emotion-specific processors. Experiments on the Weibo and Twitter datasets report accuracy of 0.935 and 0.934 respectively, outperforming ten baselines, and ablations indicate that each module contributes to the reported performance.

Significance. The knowledge-augmentation direction—using BLIP2 captions to capture image semantics and GPT-4/web evidence to break information silos—is a reasonable and potentially impactful idea, and the paper provides a modular architecture with extensive experiments. However, the central claim that emotion guidance via fine-grained emotional types improves detection is not established: the emotion reasoning loss is trained against the veracity label, the "emotional types" are undefined latent categories, and the main results lack statistical significance testing. If the emotion component can be reframed with proper emotion supervision or reinterpreted as a gated mixture-of-experts, the architecture may still be valuable; as written, the causal role of emotion modeling is unsupported.

major comments (4)
  1. [§3.3, Eq. (7)] The emotion reasoning loss L_emo uses the authenticity label y as the target for the binary emotion classifier, under the explicit assumption "negative emotion implies fake, positive emotion implies real." Therefore L_emo is a second veracity classifier operating on M_e, not an emotion signal. The ablation "w/o ER" in Table 4 consequently measures the effect of adding an extra classification branch with shared features, and the reported gains cannot be attributed to "fine-grained modeling of emotional types" as claimed. This is a load-bearing issue for the paper's second contribution.
  2. [§3.4, Eq. (9)] The x=5 "emotional types" are never defined, and no emotion labels are provided for them. The gating weights a are trained solely through the final fake news classification loss, so the balanced learning module may simply learn a latent mixture-of-experts. The claim that the model captures "inter-class differences between different emotional types of news" is therefore unsupported. The authors should either define an emotion taxonomy and provide emotion supervision, or empirically analyze what the gate partitions (e.g., via weight visualization or clustering of the gating outputs).
  3. [Table 1 and §4.1.3] The main performance comparison reports single runs without standard deviations or significance tests. Since the improvements over strong baselines such as HMCAN and CCGN on Twitter are modest (e.g., accuracy 0.934 vs. 0.906), and hyperparameters γ and λ are tuned separately per dataset, the claimed superiority is not statistically established. The authors should report multiple seeds and perform significance tests (e.g., paired bootstrap or a t-test over runs).
  4. [§4.4.4] The paper interprets the dataset-specific optimal λ values (0.20 for Weibo, 0.75 for Twitter) as evidence that "the correlation between emotional tendency and news authenticity is stronger on Twitter than on Weibo." This is not a valid inference: λ is a scalar loss weight, not a measured correlation, and its optimal value depends on the loss scale and optimization dynamics. This statement should be removed or replaced with a direct measurement of the emotion–veracity association on the data using a validated emotion classifier.
minor comments (7)
  1. [§3.3, Eq. (7)] The notation is confusing: the prediction is written ŷ_e but the text says "y_e represents the predicted binary emotion classification result," and the loss uses y as the target. Please use consistent notation, e.g., ŷ_e for the prediction and y_e for the target when it is available.
  2. [§3.2] The text says "each consisting of two Transformer [7] encoders," but reference [7] is the BERT paper; the original Transformer architecture (Vaswani et al., 2017) should be cited instead.
  3. [Figure 2 caption] The caption says "It contains of four components" and should be "It consists of four components."
  4. [§4.1.3] The number of emotional types x=5 is introduced without any definition or examples of the five categories. Please provide the emotion taxonomy or cite a standard emotion classification scheme.
  5. [General] No code or data availability statement is included; given the use of GPT-4, web search, and BLIP2, the evidence and caption generation pipeline is not reproducible from the text alone. A detailed prompt template and search procedure, or a representative sample of generated evidence, would improve reproducibility.
  6. [Table 2] The "Only Text + Evidence" and "Only Image + Caption" rows are not clean modality ablations because they combine the unimodal feature with the knowledge source but omit the other modality entirely; please clarify whether the other modality's features or CLIP embeddings are used in these settings.
  7. [Eq. (9)] The notation a = [a_1, ..., a_x] and then F = Σ a_x m_x uses a_x both for the vector and its elements; use a_i for the scalar components to avoid ambiguity.

Circularity Check

1 steps flagged · score 6.0 of 10

The emotion reasoning loss is trained on the veracity label itself, so the w/o ER ablation measures an extra veracity classifier rather than emotion modeling; the knowledge-augmentation and fusion contributions are otherwise independent.

  1. self definitional [Section 3.3, Eq. (7)]
    "Considering that fake news tends to exhibit negative emotions, while real news tends to express positive emotions, we coarsely construct an auxiliary task. We perform a binary classification task on the emotional features of the news, M_e, to distinguish between positive and negative emotions. Furthermore, we assume that news with negative emotions is fake, while news with positive emotions is real, thereby constructing the emotion reasoning loss L_emo: ... L_emo = sum -[y log(y_hat_e)+(1-y) log(1-y_hat_e)] ..."

    In Eq. (7), the target of the emotion classifier is y, the news authenticity label, under the assumption that negative emotion implies fake and positive emotion implies real. Therefore L_emo is, by construction, a second binary veracity classifier applied to M_e, not an emotion classifier with independent labels. The w/o ER ablation in Table 4 removes this extra veracity-classification branch; its accuracy gain is consistent with adding an auxiliary classifier on the same label and cannot be attributed to 'fine-grained modeling of the relationship between emotional types and authenticity'. Section 4.4.4 further interprets lambda as measuring the correlation between emotional tendency and news authenticity, but the only measured quantity is the weight on this label-prediction branch.

full rationale

The central derivation is mostly empirical and self-contained: knowledge augmentation via captions and evidence, CLIP-enhanced fusion, and balanced learning are evaluated against external baselines and ablated components that are not defined in terms of the final label. Those parts show no circularity. The one load-bearing circular step is the emotion reasoning auxiliary task: Eq. (7) defines the emotion label as the authenticity label, so the w/o ER ablation primarily demonstrates the benefit of an additional veracity classification branch rather than emotion modeling. This partial reduction by construction justifies a score of 6 rather than 0, but because the rest of KEN's contributions have independent content and no self-citation chain is load-bearing, a higher score would be disproportionate.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on four groups of assumptions: the emotion-veracity correlation used to build L_emo, which is circular because it reuses the authenticity label; the existence of separable latent emotional types that justify five expert processors; the reliability of GPT-4/web evidence; and the standard benchmark splits. Free parameters gamma, lambda, k, and x are tuned per dataset or by sweeps. The only invented entity is the set of five latent emotional categories, which has no external validation.

free parameters (4)
  • gamma (emotion modulation parameter) = 0.7 (Weibo), 0.3 (Twitter)
    Balances text vs image emotion features in Eq (6); tuned per dataset in Section 4.4.3.
  • lambda (loss balance parameter) = 0.2 (Weibo), 0.75 (Twitter)
    Weights the auxiliary emotion reasoning loss in Eq (11); tuned per dataset in Section 4.4.4.
  • number of emotion experts k = 3
    Number of averaged emotion analysis experts in Eq (5); selected by parameter sweep in Section 4.4.1.
  • number of emotional types x = 5
    Number of emotion-domain processors in Eq (8); selected by parameter sweep in Section 4.4.2.
assumptions (4)
  • domain assumption Fake news tends to express negative emotions and real news positive emotions.
    Used to construct the emotion reasoning loss L_emo, where the authenticity label y serves as the emotion label (Eq 7).
  • domain assumption There exist discrete emotional types with separable feature distributions that justify routing news to emotion-specific processors.
    Underpins balanced learning with x=5 processors (Eqs 8-9); no emotion labels or external validation are provided.
  • domain assumption Evidence retrieved from GPT-4 and web search is relevant and reliable enough to augment the short news text.
    The evidence is generated by a proprietary pipeline (Section 3.1), and only co-attention denoising is used to filter it.
  • domain assumption The standard pre-processing and data splits of the Weibo and Twitter benchmarks are appropriate and do not introduce leakage.
    Stated in Section 4.1.1 without further verification.
invented entities (1)
  • Five latent emotional categories (Type 1..Type x)
    purpose: Balanced learning maps aggregated news features into emotion-specific processors via a gating network (Eqs 8-9).
    The categories are learned implicitly; no emotion annotations, taxonomy, or validation shows they correspond to real emotional types.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KEN: Knowledge Augmentation and Emotion Guidance Network for Multimodal Fake News Detection." pith.science (2026). https://pith.science/paper/FPHVXILM

@misc{pith2026250709647,
  author       = {Pith},
  title        = {Pith review of: KEN: Knowledge Augmentation and Emotion Guidance Network for Multimodal Fake News Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FPHVXILM}},
  note         = {Machine review of arXiv:2507.09647}
}
read the original abstract

In recent years, the rampant spread of misinformation on social media has made accurate detection of multimodal fake news a critical research focus. However, previous research has not adequately understood the semantics of images, and models struggle to discern news authenticity with limited textual information. Meanwhile, treating all emotional types of news uniformly without tailored approaches further leads to performance degradation. Therefore, we propose a novel Knowledge Augmentation and Emotion Guidance Network (KEN). On the one hand, we effectively leverage LVLM's powerful semantic understanding and extensive world knowledge. For images, the generated captions provide a comprehensive understanding of image content and scenes, while for text, the retrieved evidence helps break the information silos caused by the closed and limited text and context. On the other hand, we consider inter-class differences between different emotional types of news through balanced learning, achieving fine-grained modeling of the relationship between emotional types and authenticity. Extensive experiments on two real-world datasets demonstrate the superiority of our KEN.

Figures

Figures reproduced from arXiv: 2507.09647 by the authors.

Figure 1
Figure 1. Advantages of our proposed model. (a) The gener [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed KEN. It contains of four components: Feature Representations, Knowledge Augmentation, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The experimental results indicate that the detection per￾formance exhibits a trend of first increasing and then decreasing as 𝑘 changes. When 𝑘 = 3, the KEN achieves the best performance on both datasets. When the number of experts is small, the opin￾ions of a single expert may be highly subjective or biased. Such extreme emotion analysis can adversely affect the stability and Weibo Performance Performance Twitter N… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Impact of different number of emotional types ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Impact of emotion importance adjustment between [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Impact of adjusting the loss balance parameter [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Maximizing Mutual Information Between Prompt and Response Improves LLM Performance With No Additional Data

    cs.LG 2026-03 unverdicted novelty 6.0 of 10

    MIPO constructs contrastive preference pairs from correct versus random prompts and uses DPO to maximize mutual information between prompts and responses, producing 3-40% gains on personalization and 1-18% on math tas...

  2. SEER: Semantic Enhancement and Emotional Reasoning Network for Multimodal Fake News Detection

    cs.MM 2025-07 conditional novelty 5.0 of 10

    SEER combines BLIP-2 captions, CLIP embeddings, and an expert emotion-reasoning loss to improve multimodal fake news detection on Weibo and Twitter.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages · cited by 2 Pith papers

  1. [1]

    Verify- ing multimedia use at mediaeval 2015

    Christina Boididou, Katerina Andreadou, Symeon Papadopoulos, Duc Tien Dang Nguyen, Giulia Boato, Michael Riegler, Yiannis Kompatsiaris, et al. Verify- ing multimedia use at mediaeval 2015. InMediaEval 2015, volume 1436. CEUR-WS, 2015

    work page 2015

  2. [2]

    Cross-modal ambiguity learning for multimodal fake news detection

    Yixuan Chen, Dongsheng Li, Peng Zhang, Jie Sui, Qin Lv, Lu Tun, and Li Shang. Cross-modal ambiguity learning for multimodal fake news detection. In Proceed- ings of the ACM web conference 2022 , pages 2897–2905, 2022

    work page 2022

  3. [3]

    Gin-sd: source detection in graphs with incomplete nodes via positional encoding and attentive fusion

    Le Cheng, Peican Zhu, Keke Tang, Chao Gao, and Zhen Wang. Gin-sd: source detection in graphs with incomplete nodes via positional encoding and attentive fusion. In Proceedings of the AAAI conference on artificial intelligence , volume 38, pages 55–63, 2024

    work page 2024

  4. [4]

    Efficient source detection in incomplete networks via sensor deployment and source approaching

    Le Cheng, Peican Zhu, Keke Tang, Chao Gao, and Zhen Wang. Efficient source detection in incomplete networks via sensor deployment and source approaching. IEEE Transactions on Information Forensics and Security , 2025

    work page 2025

  5. [5]

    Ccgn: consis- tency contrastive-learning graph network for multi-modal fake news detection

    Shaodong Cui, Kaibo Duan, Wen Ma, and Hiroyuki Shinnou. Ccgn: consis- tency contrastive-learning graph network for multi-modal fake news detection. Multimedia Systems, 31(2):119, 2025

    work page 2025

  6. [6]

    Approaches to identify fake news: a systematic literature review

    Dylan De Beer and Machdel Matthee. Approaches to identify fake news: a systematic literature review. Integrated Science in Digital Age 2020 , pages 13–22, 2021

    work page 2020

  7. [7]

    Bert: Pre- training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre- training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    arXiv 2018

  8. [8]

    Kan: Knowledge-aware attention network for fake news detection

    Yaqian Dun, Kefei Tu, Chen Chen, Chunyan Hou, and Xiaojie Yuan. Kan: Knowledge-aware attention network for fake news detection. In Proceedings of the AAAI conference on artificial intelligence , volume 35, pages 81–89, 2021

    work page 2021

Show all 43 references
  1. [9]

    Fake- flow: Fake news detection by modeling the flow of affective information

    Bilal Ghanem, Simone Paolo Ponzetto, Paolo Rosso, and Francisco Rangel. Fake- flow: Fake news detection by modeling the flow of affective information. arXiv preprint arXiv:2101.09810, 2021

    2021 arXiv

  2. [10]

    Leveraging emotional signals for credibility detection

    Anastasia Giachanou, Paolo Rosso, and Fabio Crestani. Leveraging emotional signals for credibility detection. In Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval , pages 877–880, 2019

    2019

  3. [11]

    Hypergraph attacks via injecting homogeneous nodes into elite hyperedges

    Meixia He, Peican Zhu, Keke Tang, and Yangming Guo. Hypergraph attacks via injecting homogeneous nodes into elite hyperedges. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 282–290, 2025

    2025

  4. [12]

    Multimodal fake news detection through data augmentation-based contrastive learning

    Jiaheng Hua, Xiaodong Cui, Xianghua Li, Keke Tang, and Peican Zhu. Multimodal fake news detection through data augmentation-based contrastive learning. Ap- plied Soft Computing, 136:110125, 2023

    2023

  5. [13]

    Multimodal fusion with recurrent neural networks for rumor detection on microblogs

    Zhiwei Jin, Juan Cao, Han Guo, Yongdong Zhang, and Jiebo Luo. Multimodal fusion with recurrent neural networks for rumor detection on microblogs. In Proceedings of the 25th ACM international conference on Multimedia , pages 795– 816, 2017

    2017

  6. [14]

    Mvae: Multimodal variational autoencoder for fake news detection

    Dhruv Khattar, Jaipal Singh Goud, Manish Gupta, and Vasudeva Varma. Mvae: Multimodal variational autoencoder for fake news detection. In The world wide web conference, pages 2915–2921, 2019

    2019

  7. [15]

    Disinformation on the web: Impact, characteristics, and detection of wikipedia hoaxes

    Srijan Kumar, Robert West, and Jure Leskovec. Disinformation on the web: Impact, characteristics, and detection of wikipedia hoaxes. In Proceedings of the 25th international conference on World Wide Web , pages 591–602, 2016

    2016

  8. [16]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning , pages 19730–19742. PMLR, 2023

    2023

  9. [17]

    Entity-oriented multi-modal alignment and fusion network for fake news detection

    Peiguang Li, Xian Sun, Hongfeng Yu, Yu Tian, Fanglong Yao, and Guangluan Xu. Entity-oriented multi-modal alignment and fusion network for fake news detection. IEEE Transactions on Multimedia , 24:3455–3468, 2021

    2021

  10. [18]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021

    2021

  11. [19]

    Vilbert: Pretraining task- agnostic visiolinguistic representations for vision-and-language tasks

    Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task- agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems , 32, 2019

    2019

  12. [20]

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

    Alex Munyole Luvembe, Weimin Li, Shaohua Li, Fangfang Liu, and Guiqiong Xu. Dual emotion based fake news detection: A deep attention-weight update approach. Information Processing & Management , 60(4):103354, 2023

    2023

  13. [21]

    Detecting rumors from microblogs with recurrent neural networks

    Jing Ma, Wei Gao, Prasenjit Mitra, Sejeong Kwon, Bernard J Jansen, Kam-Fai Wong, and Meeyoung Cha. Detecting rumors from microblogs with recurrent neural networks. 2016

    2016

  14. [22]

    Finding deceptive opinion spam by any stretch of the imagination

    Myle Ott, Yejin Choi, Claire Cardie, and Jeffrey T Hancock. Finding deceptive opinion spam by any stretch of the imagination. arXiv preprint arXiv:1107.4557, 2011

    2011 arXiv

  15. [23]

    Improving fake news detection by using an entity-enhanced framework to fuse diverse multimodal clues

    Peng Qi, Juan Cao, Xirong Li, Huan Liu, Qiang Sheng, Xiaoyue Mi, Qin He, Yongbiao Lv, Chenyang Guo, and Yingchao Yu. Improving fake news detection by using an entity-enhanced framework to fuse diverse multimodal clues. In Proceedings of the 29th ACM International Conference on...

    2021

  16. [24]

    Exploiting multi- domain visual information for fake news detection

    Peng Qi, Juan Cao, Tianyun Yang, Junbo Guo, and Jintao Li. Exploiting multi- domain visual information for fake news detection. In 2019 IEEE international conference on data mining (ICDM) , pages 518–527. IEEE, 2019

    2019

  17. [25]

    Hierarchical multi-modal contextual attention network for fake news detection

    Shengsheng Qian, Jinguang Wang, Jun Hu, Quan Fang, and Changsheng Xu. 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, pages 153–162, 2021

    2021

  18. [26]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , p...

    2021

  19. [27]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  20. [28]

    Spotfake: A multi-modal framework for fake news detection

    Shivangi Singhal, Rajiv Ratn Shah, Tanmoy Chakraborty, Ponnurangam Ku- maraguru, and Shin’ichi Satoh. Spotfake: A multi-modal framework for fake news detection. In 2019 IEEE fifth international conference on multimedia big data (BigMM), pages 39–47. IEEE, 2019

    2019

  21. [29]

    Visualizing data using t-sne

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research , 9(11), 2008

    2008

  22. [30]

    Positive unla- beled fake news detection via multi-modal masked transformer network

    Jinguang Wang, Shengsheng Qian, Jun Hu, and Richang Hong. Positive unla- beled fake news detection via multi-modal masked transformer network. IEEE Transactions on Multimedia, 26:234–244, 2024

    2024

  23. [31]

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

    Yaqing Wang, Fenglong Ma, Zhiwei Jin, Ye Yuan, Guangxu Xun, Kishlay Jha, Lu Su, and Jing Gao. 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 , page...

    2018

  24. [32]

    Fake news detection via knowledge-driven multimodal graph convolutional networks

    Youze Wang, Shengsheng Qian, Jun Hu, Quan Fang, and Changsheng Xu. Fake news detection via knowledge-driven multimodal graph convolutional networks. In Proceedings of the 2020 international conference on multimedia retrieval , pages 540–547, 2020

    2020

  25. [33]

    See how you read? multi-reading habits fusion reasoning for multi-modal fake news detection

    Lianwei Wu, Pusheng Liu, and Yanning Zhang. See how you read? multi-reading habits fusion reasoning for multi-modal fake news detection. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 37, pages 13736–13744, 2023

    2023

  26. [34]

    Multimodal fusion with co-attention networks for fake news detection

    Yang Wu, Pengwei Zhan, Yunjian Zhang, Liming Wang, and Zhen Xu. Multimodal fusion with co-attention networks for fake news detection. In Findings of the association for computational linguistics: ACL-IJCNLP 2021, pages 2560–2569, 2021

    2021

  27. [35]

    Gamc: an unsu- pervised method for fake news detection using graph autoencoder with masking

    Shu Yin, Peican Zhu, Lianwei Wu, Chao Gao, and Zhen Wang. Gamc: an unsu- pervised method for fake news detection using graph autoencoder with masking. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 347–355, 2024

    2024

  28. [36]

    A convolutional approach for misinformation identification

    Feng Yu, Qiang Liu, Shu Wu, Liang Wang, Tieniu Tan, et al. A convolutional approach for misinformation identification. In IJCAI, pages 3901–3907, 2017

    2017

  29. [37]

    Multi-modal knowledge-aware event memory network for social media rumor detection

    Huaiwen Zhang, Quan Fang, Shengsheng Qian, and Changsheng Xu. Multi-modal knowledge-aware event memory network for social media rumor detection. In Proceedings of the 27th ACM international conference on multimedia , pages 1942– 1951, 2019

    1942

  30. [38]

    Mining dual emotion for fake news detection

    Xueyao Zhang, Juan Cao, Xirong Li, Qiang Sheng, Lei Zhong, and Kai Shu. Mining dual emotion for fake news detection. In Proceedings of the web conference 2021 , pages 3465–3476, 2021

    2021

  31. [39]

    : Similarity-aware multi-modal fake news detection

    Xinyi Zhou, Jindi Wu, and Reza Zafarani. : Similarity-aware multi-modal fake news detection. In Pacific-Asia Conference on knowledge discovery and data mining, pages 354–367. Springer, 2020

    2020

  32. [40]

    Multi-modal fake news detection on social media via multi-grained information fusion

    Yangming Zhou, Yuzhou Yang, Qichao Ying, Zhenxing Qian, and Xinpeng Zhang. Multi-modal fake news detection on social media via multi-grained information fusion. In Proceedings of the 2023 ACM International Conference on Multimedia Retrieval, pages 343–352, 2023

    2023

  33. [41]

    Multimodal fake news detection via clip-guided learning

    Yangming Zhou, Yuzhou Yang, Qichao Ying, Zhenxing Qian, and Xinpeng Zhang. Multimodal fake news detection via clip-guided learning. In 2023 IEEE Inter- national Conference on Multimedia and Expo (ICME) , pages 2825–2830. IEEE, 2023

    2023

  34. [42]

    Multimodal fake news detection through intra-modality feature aggregation and inter-modality semantic fusion

    Peican Zhu, Jiaheng Hua, Keke Tang, Jiwei Tian, Jiwei Xu, and Xiaodong Cui. Multimodal fake news detection through intra-modality feature aggregation and inter-modality semantic fusion. Complex & Intelligent Systems , 10(4):5851–5863, 2024

    2024

  35. [43]

    A general black-box adversarial attack on graph-based fake news detectors

    Peican Zhu, Zechen Pan, Yang Liu, Jiwei Tian, Keke Tang, and Zhen Wang. A general black-box adversarial attack on graph-based fake news detectors. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 568–576, 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.