Pith. sign in

REVIEW 3 major objections 4 minor 60 references

Dataset of News Articles with Provenance Metadata for Media Relevance Assessment

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper introduces the first news dataset whose images carry provenance metadata, defines location and date relevance tasks over that metadata, and reports baseline LLM results showing location relevance is tractable but date relevance…

desk verdict A genuinely useful first benchmark, but the DTOR numbers are too entangled with label noise to support the paper's main claim. read the letter →

arxiv 2506.09847 v1 pith:S2FXT7L4 submitted 2025-06-11 cs.CL cs.AIcs.CVcs.CY

classification cs.CLcs.AIcs.CVcs.CY
keywords provenancemetadatanewsimagerelevanceout-of-contextimageryC2PAlocationoforigindateandtimemultimodalmisinformationLLMbaseline
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's central assertion is that provenance metadata—the recorded location and date/time of an image's capture—can catch a form of news-media manipulation that pixel-level semantic checks miss: images that match the story's topic but were taken at an irrelevant place or time. To make this testable, the authors introduce the News Media Provenance Dataset, 637 news articles whose main images carry simulated C2PA provenance metadata labeled relevant or not relevant to the article. They define two tasks, location of origin relevance (LOR) and date and time of origin relevance (DTOR), and report zero-shot baseline accuracy for six large language models, finding LOR accuracies of 64–81 percent but DTOR accuracies of 42–58 percent. If the dataset and task framing hold, it gives researchers and newsrooms a benchmark for asking not 'does this image fit the story' but 'was this image actually taken where and when the story says.'

What carries the argument

The carrying object is the News Media Provenance Dataset itself: a collection of 637 article-image pairs with C2PA-embedded provenance metadata, where each image appears with human-annotated relevant metadata and, in many cases, LLM-generated non-relevant alternatives. The paper's two proposed tasks, LOR and DTOR, isolate the relevance of the stated capture location and capture time to the article text, so a model must compare metadata against the article's described events rather than judge semantic fit between image pixels and text. The C2PA metadata wrapper is what makes the simulation concrete, since provenance is not yet widely adopted by the news outlets the articles were scraped from.

What would settle it

Take a random sample of the dataset's not-relevant provenance entries created by ChatGPT-4o and have independent annotators judge whether the generated place or time could plausibly match the article; if a substantial share are accidentally relevant, the negative labels are invalid. Separately, re-annotate the date-of-origin labels with consensus across more annotators and recompute the six DTOR accuracies; if scores move materially, the reported 'DTOR is hard' conclusion was driven by label noise.

Watch

Extended reading notes

Core claim

The paper claims to be the first to offer a news dataset whose images carry provenance metadata and to use that metadata as the object of relevance assessment. The authors simulate provenance by having four human annotators infer the likely location and date of capture for each article's main image, embedding the human-chosen values as relevant metadata via C2PA, and having ChatGPT-4o generate deliberately unrelated locations and times as not-relevant metadata. Their baseline results show that off-the-shelf LLMs can reason about LOR reasonably well, with ChatGPT-4o at 81 percent, but DTOR remains hard, with the best model at 58 percent and three smaller models at or below 50 percent; qualitative analysis attributes the shortfall to LLMs' weak temporal reasoning, such as conflating the article time frame with the analysis time frame.

Load-bearing premise

The load-bearing premise is that the human annotations and LLM-generated alternatives together define 'relevant' and 'not relevant' provenance correctly, and that premise is shaky for dates, where four annotators agreed only 56 percent of the time, and for negative examples, which were never human-checked.

Editorial extensions

If this is right

  • LOR becomes a usable zero-shot screening signal: a newsroom or platform with access to C2PA metadata could flag likely location-mismatched images with off-the-shelf LLMs at 64–81 percent accuracy, without any task-specific training.
  • DTOR, by contrast, is an open problem: current LLMs are at or near chance, so any future provenance-aware system needs specialized temporal reasoning rather than a prompt-only fix.
  • The dataset supplies the first common yardstick for provenance-based relevance assessment, allowing subsequent methods to report numbers comparable to the six baselines rather than inventing ad-hoc evaluations.
  • Because the LOR and DTOR tasks deliberately ignore pixels, they measure the value of provenance metadata alone, complementing semantic out-of-context datasets like COSMOS rather than replacing them.
  • If C2PA adoption among outlets grows, the same two-task formulation transfers directly from simulated metadata to real provenance fields without changing the evaluation protocol.

Reading between the lines

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

  • The low annotator agreement on dates (56 percent) suggests DTOR accuracy numbers may be partly a measurement of label noise; consensus-based re-annotation of the date labels would likely tighten or shift every baseline.
  • The 'not relevant' provenance values were generated by ChatGPT-4o and never human-checked, so some may accidentally coincide with the article's actual place or time; a human audit of negative labels would test whether reported accuracies are fair.
  • If proven correct, the same relevance logic could be applied outside newsrooms, for example to social-media resharing where camera-applied C2PA metadata accompanies photos, making provenance relevance a general misinformation signal.
  • The paper's own limitation about illustrative images suggests a boundary condition: the approach assumes a tight match between metadata and article event, so a broader system would need to decide when strict alignment is required before applying LOR and DTOR.
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 / 4 minor

Summary. The paper introduces the News Media Provenance Dataset, a collection of 637 news articles with simulated C2PA provenance metadata attached to their main images. Two new tasks are defined: Location of Origin Relevance (LOR) and Date and Time of Origin Relevance (DTOR), each framed as a binary classification of whether the metadata is relevant to the article text. The authors report zero-shot baseline accuracies for six LLMs, finding LOR accuracy between 64 and 81 percent and DTOR accuracy between 42 and 58 percent, and conclude that current models handle location relevance reasonably well but struggle with temporal relevance.

Significance. If the dataset and labels are valid, this would be a useful first benchmark for provenance-based media relevance assessment, a genuinely understudied problem in the out-of-context misinformation literature. The paper is transparent about its construction pipeline, releases the dataset and evaluation code, and provides qualitative error analyses. The LOR/DTOR task formulation is a natural and valuable extension of semantic image-text relevance. However, the empirical conclusions rest on label quality that is not established: the DTOR annotations have low inter-annotator agreement, and the negative labels are generated by an LLM without human verification. The resource could become a solid contribution after label-quality fixes, but in its current form the central DTOR finding is not reliably supported.

major comments (3)
  1. [Section 3.1.3 and Table 1] The DTOR ground truth is too noisy to support the paper's main empirical claim. On the five shared articles, annotator accuracy for date of origin is only 56% (with a plus/minus one unit buffer), and the remaining 195 articles are each labeled by a single annotator with no adjudication. Since Table 1 reports DTOR accuracies between 42% and 58%, the observed differences among models and the overall 'DTOR hinders' conclusion are at or below the measured label-noise level. The LOR>DTOR gap may be an artifact of the relative unreliability of the date labels rather than a genuine difference in model capability. The authors should either re-annotate the date-of-origin labels with multiple annotators and compute agreement on the full dataset, or report results only after adjudication.
  2. [Section 3.2 and Section 4] The negative provenance labels are generated by ChatGPT-4o with a prompt requesting 'absolutely unrelated' places and times, and there is no human verification of these labels. A generated date or place can be accidentally relevant for articles reporting on events with broad temporal or geographical scope, as the paper itself acknowledges in Section 7 for natural disasters. Because ChatGPT-4o is also one of the six models evaluated in Section 4, the evaluation contains an unexamined distributional dependence: part of the test set is produced by the same model family under test. The authors should either have the generated negatives checked by human annotators against a clear notion of relevance, or provide evidence that ChatGPT-4o-generated alternatives are reliably non-relevant.
  3. [Sections 3.3 and 5.1] The dataset size and task statistics are under-specified in a way that affects interpretation of the baseline numbers. The paper states there are 637 news articles, but Section 3.2 says each article yields up to three additional data points (one partially irrelevant, two both-irrelevant). It is unclear how many total LOR and DTOR examples were used for the accuracy results in Table 1, and whether the reported accuracies are per data point or per article. This matters because the article-level statistics are given separately, and small sample sizes for feature-level results would make the reported differences (e.g., 57% vs 56% for ChatGPT-4o and DeepSeek V3) indistinguishable from noise. The authors should report the exact number of evaluation examples per condition and include confidence intervals or significance tests.
minor comments (4)
  1. [Section 3.3] The text says 'The top-10 domains ... are shown in Figure 3,' but Figure 3 is the token-length distribution; the domain distribution appears to be Figure 5. Please correct the cross-reference.
  2. [Section 6] The word 'measuing' in the final paragraph should be 'measuring.'
  3. [Abstract and Section 1] The phrase 'we introduce News Media Provenance Dataset' is missing a definite article; it should be 'the News Media Provenance Dataset' for consistency with the rest of the paper.
  4. [Section 3.1.1] There is a repeated 'the' in 'from the the Webz.io News Dataset Repository.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dataset and baselines are constructed and measured independently of the paper's claims.

full rationale

The paper's claimed contributions are a new dataset, two task formulations, and zero-shot baselines. There is no fitted parameter later renamed as a prediction, no uniqueness theorem imported from the authors' prior work, and no ansatz smuggled in via citation. Positive provenance labels are produced by human annotators from article context (Section 3.1.2), and negative labels are produced by ChatGPT-4o (Section 3.2); although the same model family appears in the baseline table, the generation prompt asks for values 'absolutely unrelated' to the original place/time, not for the article-relevance label the model is later asked to predict, so the evaluation is not equivalent to the generation step by construction. The 56% date-annotation agreement and the absence of human validation for generated negatives are threats to benchmark validity, not circularity in the derivation chain. The self-citation of Peterka and Bohacek (2025) motivates the task formulation but is not load-bearing: the dataset construction, tasks, and all six baseline evaluations are self-contained and open-sourced. No equation or definition in the paper reduces a reported result to its own input.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

No fitted free parameters or invented physical entities appear in this benchmark paper. The quantitative conclusions rest on two label-validity assumptions: human annotations define relevant provenance, and LLM-generated alternatives define non-relevant provenance. Both are explicit but only the former is partially measured.

free parameters (1)
  • Date evaluation buffer (±1 unit) = ±1 month (or year)
    Human date labels are graded with a ±1 buffer in units of the provided granularity (Section 3.1.3, Appendix A). This hand-set tolerance changes the measured annotator reliability (56%) and the effective difficulty of DTOR.
assumptions (3)
  • domain assumption Human annotators' inferred location and month/year for an image constitute a valid ground-truth 'relevant' provenance.
    The dataset labels relevant provenance from four Prolific annotators reading the article and image caption; reliability is 80% for location and 56% for date (Section 3.1.3), so this assumption is only partially validated.
  • ad hoc to paper ChatGPT-4o-generated 'absolutely unrelated' locations and dates are valid 'not relevant' labels.
    Negative provenance is generated by prompting an LLM to produce unrelated places and times (Section 3.2, Appendix B) and is not human-validated.
  • domain assumption C2PA provenance metadata, when embedded, accurately represents the image's capture circumstances for evaluation.
    The paper simulates provenance by embedding annotator/LLM metadata with C2PA (Sections 2.3 and 3); real-world C2PA adoption is currently limited, which the authors acknowledge in Section 7.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dataset of News Articles with Provenance Metadata for Media Relevance Assessment." pith.science (2026). https://pith.science/paper/S2FXT7L4

@misc{pith2026250609847,
  author       = {Pith},
  title        = {Pith review of: Dataset of News Articles with Provenance Metadata for Media Relevance Assessment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S2FXT7L4}},
  note         = {Machine review of arXiv:2506.09847}
}
read the original abstract

Out-of-context and misattributed imagery is the leading form of media manipulation in today's misinformation and disinformation landscape. The existing methods attempting to detect this practice often only consider whether the semantics of the imagery corresponds to the text narrative, missing manipulation so long as the depicted objects or scenes somewhat correspond to the narrative at hand. To tackle this, we introduce News Media Provenance Dataset, a dataset of news articles with provenance-tagged images. We formulate two tasks on this dataset, location of origin relevance (LOR) and date and time of origin relevance (DTOR), and present baseline results on six large language models (LLMs). We identify that, while the zero-shot performance on LOR is promising, the performance on DTOR hinders, leaving room for specialized architectures and future work.

Figures

Figures reproduced from arXiv: 2506.09847 by the authors.

Figure 1
Figure 1. Representative example of a news article from the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Examples of images from the News Media Provenance Dataset used to evaluate annotator reliability. All four annotators provided the location and date of origin for each image, with their accuracy indicated on the right. The article at the top is sourced from CBC and the article at the bottom is sourced from Forbes. nance metadata enables, it seems highly desirable for relevance assessment of imagery in news arti￾cles… view at source ↗
Figure 3
Figure 3. Distribution of the title, body, and image caption token length in the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Distribution of source domains in the News Media Provenance Dataset, showing the top 10 do￾mains. The top-10 domains by absolute article count are yahoo, dailymail, cbsnews, foxnews, euronews, aljazeera, cbc, forbes, nbcmiami, and usatoday, as shown in [PITH_FULL_IMAG…
Figure 6
Figure 6. Figure 6: Screenshot of the Argilla annotation tool, focused on an article body. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Screenshot of the Argilla annotation tool, focused on an image and its caption. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Screenshot of the Argilla annotation tool with the instructions window open. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 28 canonical work pages

  1. [1]

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. 2024. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219

  2. [2]

    AndyTheFactory. 2023. https://github.com/AndyTheFactory/newspaper4k Newspaper4k : A rticle scraping & curation

  3. [3]

    Shivangi Aneja, Chris Bregler, and Matthias Nie ner. 2021. Cosmos: Catching out-of-context misinformation with self-supervised learning. arXiv preprint arXiv:2101.06278

  4. [4]

    Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Lluis Castrejon, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, et al. 2024. Imagen 3. arXiv preprint arXiv:2408.07009

  5. [5]

    Fengxiang Bie, Yibo Yang, Zhongzhu Zhou, Adam Ghanem, Minjia Zhang, Zhewei Yao, Xiaoxia Wu, Connor Holmes, Pareesa Golnari, David A Clifton, et al. 2024. Renaissance: A survey into ai text-to-image generation in the era of large model. IEEE Transactions on Pattern Analysis and Machine Intelligence

  6. [6]

    Steven Bird. 2006. Nltk: the natural language toolkit. In Proceedings of the COLING/ACL 2006 interactive presentation sessions, pages 69--72

  7. [7]

    Matyas Bohacek, Michal Bravansky, Filip Trhl \' k, and V \'a clav Moravec. 2023. Czech-ing the news: Article trustworthiness dataset for czech. In Proceedings of the 13th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, pages 96--109

  8. [8]

    Gullal S Cheema, Sherzod Hakimov, Eric M \"u ller-Budack, Christian Otto, John A Bateman, and Ralph Ewerth. 2023. Understanding image-text relations and news values for multimodal news analysis. Frontiers in artificial intelligence, 6:1125533

Show all 60 references
  1. [9]

    Sanyuan Chen, Shujie Liu, Long Zhou, Yanqing Liu, Xu Tan, Jinyu Li, Sheng Zhao, Yao Qian, and Furu Wei. 2024. VALL-E 2 : N eural codec language models are human parity zero-shot text to speech synthesizers. arXiv preprint arXiv:2406.05370

  2. [10]

    Coalition for Content Provenance and Authenticity (C2PA) . 2023. https://c2pa.org/specifications/specifications/1.0/security/_attachments/Initial_Adoption_Assessment.pdf Harms, Misuse, and Abuse: Initial Adoption Assessment

  3. [11]

    Florinel-Alin Croitoru, Andrei-Iulian Hiji, Vlad Hondru, Nicolae Catalin Ristea, Paul Irofti, Marius Popescu, Cristian Rusu, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. 2024. Deepfake media generation and detection in the generative ai era: A survey and outlook. ...

  4. [12]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

  5. [13]

    Nicholas Dufour, Arkanath Pathak, Pouya Samangouei, Nikki Hariri, Shashi Deshetti, Andrew Dudfield, Christopher Guess, Pablo Hern \'a ndez Escayola, Bobby Tran, Mevan Babakar, et al. 2024. AMMeBa : A large-scale survey and dataset of media-based misinformation in-the-wild. arX...

  6. [14]

    Sefik Emre Eskimez, Xiaofei Wang, Manthan Thakker, Canrun Li, Chung-Hsien Tsai, Zhen Xiao, Hemin Yang, Zirun Zhu, Min Tang, Xu Tan, et al. 2024. E2 TTS : E mbarrassingly easy fully non-autoregressive zero-shot tts. In 2024 IEEE Spoken Language Technology Workshop (SLT), pages ...

  7. [15]

    Hany Farid. 2022. Creating, using, misusing, and detecting deep fakes. Journal of Online Trust and Safety, 1(4)

  8. [16]

    Lisa Fazio. 2020. Out-of-context photos are a powerful low-tech form of misinformation. The Conversation, 14(1)

  9. [17]

    Kiran Garimella and Dean Eckles. 2020. Images and misinformation in political groups: Evidence from whatsapp in india. arXiv preprint arXiv:2005.09784

  10. [18]

    Maur \' cio Gruppi, Benjamin D Horne, and Sibel Adal . 2021. Nela-gt-2020: A large multi-labelled news dataset for the study of misinformation in news articles. arXiv preprint arXiv:2102.04567

  11. [19]

    Max Grusky, Mor Naaman, and Yoav Artzi. 2018. Newsroom: A dataset of 1.3 million summaries with diverse extractive strategies. arXiv preprint arXiv:1804.11283

  12. [20]

    Antonio Gulli. 2005. The anatomy of a news search engine. In Special interest tracks and posters of the 14th international conference on World Wide Web, pages 880--881

  13. [21]

    Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. Advances in neural information processing systems, 28

  14. [22]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

  15. [23]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

  16. [24]

    Xiang Jiang and Markus Dreyer. 2024. Ccsum: A large-scale and high-quality dataset for abstractive news summarization. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: L...

  17. [25]

    Mateusz ajszczak, Guillermo C \'a mbara, Yang Li, Fatih Beyhan, Arent Van Korlaar, Fan Yang, Arnaud Joly, \'A lvaro Mart \' n-Cortinas, Ammar Abbas, Adam Michalski, et al. 2024. BASE TTS : L essons from building a billion-parameter text-to-speech model on 100k hours of data. a...

  18. [26]

    Ken Lang. 1995. Newsweeder: Learning to filter netnews. In Machine learning proceedings 1995, pages 331--339. Elsevier

  19. [27]

    Eunhye Lee, Jeongmu Kim, Jisu Kim, and Tae Hyun Kim. 2021. Restore from restored: Single-image inpainting. arXiv preprint arXiv:2102.08078

  20. [28]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. BART : D enoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461

  21. [29]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437

  22. [30]

    Anji Liu, Mathias Niepert, and Guy Van den Broeck. 2023. Image inpainting via tractable steering of diffusion models. arXiv preprint arXiv:2401.03349

  23. [31]

    Shayne Longpre, Robert Mahari, Naana Obeng-Marnu, William Brannon, Tobin South, Jad Kabbara, and Sandy Pentland. 2024. Data authenticity, consent, and provenance for ai are all broken: What will it take to fix them?

  24. [32]

    Hieu-Thi Luong and Junichi Yamagishi. 2020. Nautilus: a versatile voice cloning system. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2967--2981

  25. [33]

    Rishabh Misra. 2022. News category dataset. arXiv preprint arXiv:2209.11429

  26. [34]

    News Literacy Project . 2025. https://newslit.org/tips-tools/covid-19-video-out-of-context/ Covid-19 video taken out of context . Accessed: 2025-02-23

  27. [35]

    Thanh Thi Nguyen, Quoc Viet Hung Nguyen, Dung Tien Nguyen, Duc Thanh Nguyen, Thien Huynh-The, Saeid Nahavandi, Thanh Tam Nguyen, Quoc-Viet Pham, and Cuong M Nguyen. 2022. Deep learning for deepfakes creation and detection: A survey. Computer Vision and Image Understanding, 223:103525

  28. [36]

    Rubungo Andre Niyongabo, Hong Qu, Julia Kreutzer, and Li Huang. 2020. Kinnews and kirnews: Benchmarking cross-lingual text classification for kinyarwanda and kirundi. arXiv preprint arXiv:2010.12174

  29. [37]

    Lucas Ou-Yang. 2013. Newspaper3k : A rticle scraping & curation. Newspaper3k: Article Scraping & Curation-Newspaper 0.0. 2 Documentation

  30. [38]

    Gan Pei, Jiangning Zhang, Menghan Hu, Zhenyu Zhang, Chengjie Wang, Yunsheng Wu, Guangtao Zhai, Jian Yang, Chunhua Shen, and Dacheng Tao. 2024. Deepfake generation and detection: A benchmark and survey. arXiv preprint arXiv:2403.17881

  31. [39]

    Tomas Peterka and Matyas Bohacek. 2025. Large language models and provenance metadata for determining the relevance of images and videos in news stories. arXiv preprint arXiv:2502.09689

  32. [40]

    Alina Petukhova and Nuno Fachada. 2023. Mn-ds: A multilabeled news dataset for news articles hierarchical classification. Data, 8(5):74

  33. [41]

    Zengyi Qin, Wenliang Zhao, Xumin Yu, and Xin Sun. 2023. Openvoice: Versatile instant voice cloning. arXiv preprint arXiv:2312.01479

  34. [42]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821--8831. Pmlr

  35. [43]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684--10695

  36. [44]

    Leonard Rosenthol. 2022. C2pa: the world’s first industry standard for content provenance (conference presentation). In Applications of Digital Image Processing XLV, volume 12226, page 122260P. SPIE

  37. [45]

    Cuihua Shen, Mona Kasra, and James O'Brien. 2021. This photograph has been altered: Testing the effectiveness of image forensic labeling on news image credibility. arXiv preprint arXiv:2101.07951

  38. [46]

    Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. 2022. Make-A-Video : T ext-to-video generation without text-video data. arXiv preprint arXiv:2209.14792

  39. [47]

    Valeriya Slovikovskaya. 2019. Transfer learning from transformers to fake news challenge stance detection (fnc-1) task. arXiv preprint arXiv:1910.14353

  40. [48]

    Georgii Stanishevskii, Jakub Steczkiewicz, Tomasz Szczepanik, S awomir Tadeja, Jacek Tabor, and Przemys aw Spurek. 2024. Implicitdeepfake: Plausible face-swapping through implicit deepfake generation using nerf and gaussian splatting. arXiv e-prints, pages arXiv--2402

  41. [49]

    Milan Straka, Nikita Mediankin, Tom Kocmi, Zden e k Z abokrtsk \`y , Vojt e ch Hude c ek, and Jan Hajic. 2018. Sumeczech: Large czech news-based summarization dataset. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)

  42. [50]

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, L \'e onard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ram \'e , et al. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118

  43. [51]

    image, tell me your story!

    Jonathan Tonglet, Marie-Francine Moens, and Iryna Gurevych. 2024. "image, tell me your story!" predicting the original meta-context of visual misinformation. arXiv preprint arXiv:2408.09939

  44. [52]

    Bing Wang, Shengsheng Wang, Changchun Li, Renchu Guan, and Ximing Li. 2024. Harmfully manipulated images matter in multimodal misinformation detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2262--2271

  45. [53]

    liar, liar pants on fire

    William Yang Wang. 2017. " liar, liar pants on fire": A new benchmark dataset for fake news detection. arXiv preprint arXiv:1705.00648

  46. [54]

    Webhose.io. 2024. https://github.com/Webhose/free-news-datasets Free news datasets

  47. [55]

    Teresa Weikmann and Sophie Lecheler. 2023. Visual disinformation in a digital age: A literature synthesis and research agenda. New Media & Society, 25(12):3696--3713

  48. [56]

    Thomas Wolf. 2020. Transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771

  49. [57]

    Yejun Yoon, Seunghyun Yoon, and Kunwoo Park. 2024. Understanding news thumbnail representativeness by counterfactual text-guided contrastive language-image pretraining. arXiv preprint arXiv:2402.11159

  50. [58]

    Shilong Zhang, Wenbo Li, Shoufa Chen, Chongjian Ge, Peize Sun, Yida Zhang, Yi Jiang, Zehuan Yuan, Binyue Peng, and Ping Luo. 2025. Flashvideo: Flowing fidelity to detail for efficient high-resolution video generation. arXiv preprint arXiv:2502.05179

  51. [59]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  52. [60]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.