Back

Faking Fake News for Real Fake News Detection: Propaganda-Loaded Training Data Generation

Kung-Hsiang HuangKathleen R. McKeownPreslav NakovYejin ChoiHeng Ji

article2023ACL86 citations

Proposes an automated framework that combines natural language inference-guided infilling with propaganda generation to create realistic synthetic disinformation that significantly improves the detection of human-written fake news.

Listen

Online disinformation poses a growing threat to public trust and social stability, yet building automated systems to detect it is hindered by a lack of high-quality training data. Manually fact-checking human-written disinformation is slow and unscalable, while standard machine-generated fake news fails to reflect how real disinformation is crafted. In reality, human-authored disinformation rarely fabricates an entire story; instead, it strategically embeds false claims within otherwise factual reporting and employs psychological manipulation through propaganda techniques.

The article demonstrates a novel framework to automatically generate realistic training data that mirrors the structure and style of human-written disinformation. Its primary objective is to evaluate whether training machine learning detectors on this propaganda-loaded synthetic data improves their accuracy in spotting actual human-written fake news.

To achieve this, the researchers developed a multi-stage approach using authentic news articles as a foundation. They first identified the most critical, central sentence in an article and replaced it with a plausible but false alternative. To ensure the new sentence was truly disinformative and not logically implied by the original context, they applied a natural language inference reward mechanism during generation. Next, they injected two prevalent propaganda techniques: appeal to authority, which fabricates quotes from domain experts, and loaded language, which adds sensationalized, emotion-triggering phrasing. The researchers used this pipeline to create a benchmark dataset containing 2,256 verified articles and evaluated several transformer-based detection models across standard public benchmarks.

The findings confirm that this generation strategy significantly enhances detection capabilities. First, detectors trained on the new dataset achieved a 3.62% to 7.69% improvement in detection performance on real-world benchmark datasets compared to models trained on existing synthetic baselines. Second, the generated articles demonstrated a substantially closer statistical similarity to real disinformation than previous fully synthetic methods. Third, human evaluation confirmed that incorporating propaganda significantly increased the plausibility of the generated articles, with nearly 30% of evaluators noting that the propaganda techniques directly influenced their perception of the story's credibility. Finally, ablation tests proved that adding both emotional and authoritative propaganda directly contributed to the overall performance gains of the detectors.

These results demonstrate that effective disinformation detection requires training models on subtle, mixed-factuality text rather than entirely fabricated articles. By capturing stylistic nuances and rhetorical manipulation, synthetic data generation becomes a viable, cost-effective substitute for labor-intensive manual curation. Furthermore, identifying specific manipulated sentences enables more explainable and auditable detection systems, reducing compliance and operational risks for platforms tasked with content moderation.

Organizations developing automated moderation or intelligence analysis tools should transition from fully synthetic training regimes to targeted, manipulation-based data pipelines. To manage ethical and dual-use risks, practitioners must implement strict governance, as the underlying generation techniques could potentially be repurposed to generate harmful disinformation. For safety, the source code for the detectors and the resulting dataset are made available, while the direct generation engine is restricted.

Certain limitations remain, warranting measured confidence in specific operating environments. Performance gains plateaued when scaling synthetic data volume, indicating an enduring gap in capturing human intent, narrative novelty, and broader deceptive styles. The framework is currently limited to English-language texts and automates only two specific propaganda techniques, while failing to detect disinformation that requires dynamic, real-time knowledge or multi-document reasoning. Future initiatives should focus on broadening language coverage, integrating additional propaganda strategies, and connecting detectors to live, external knowledge bases.

detection algorithmPropaganda-loaded training data generationFake news detectionPropaganda detection
Huang et al (2023).pdf
Cover for Faking Fake News for Real Fake News Detection: Propaganda-Loaded Training Data Generation

Topics

detection algorithm

Keywords

Propaganda-loaded training data generationFake news detectionPropaganda detectionSelf-critical sequence trainingPolitiFact dataset

Table of Contents

  • 1 Introduction
  • 2 Training Data Generation
  • 2.1 Disinformation Generation
  • 2.2 Propaganda Generation
  • 2.3 Intermediate Pre-training
  • 3 Our PROPANEWS Dataset
  • 3.1 Data Source
  • 3.2 Crowdsourcing for Data Curation
  • 4 Disinformation Detection
  • 5 Experiments
  • 5.1 Data
  • 5.2 Results and Discussion
  • 6 Related Work
  • 7 Conclusions and Future Work
  • 8 Limitations
  • 9 Ethical Statement and Broader Impact
  • Acknowledgement
  • References
  • A Distribution of Propaganda
  • B Additional Research Questions
  • C Further Analysis
  • C.1 Remaining Challenges
  • D Qualitative Examples of Generated Articles
  • E Appeal to Authority Details
  • F Intermediate Pre-training Details
  • G Benchmarking Detectors
  • H Human Validation Details
  • I Statistics about the Evaluation Datasets
  • J Detector Implementation Details
  • K Human Evaluation Details
  • L Scientific Artifacts

Knowls

  1. Knowl 1 — Generate disinformation by replacing a salient sentence in authentic news

    model/method

    The paper’s generation framework starts with an authentic news article and changes one sentence while preserving the surrounding article. An extractive summarizer assigns each sentence a score estimating its likelihood of appearing in a summary; the highest-scoring sentence is treated as salient. BART then infills that sentence’s position using the remaining article as context. To make the replacement disinformative rather than merely a paraphrase, generation is trained with a natural-language-inference (NLI) objective that discourages the original sentence from entailing the replacement, and an NLI filter discards replacements that remain entailed. The resulting articles can then be augmented with appeal-to-authority or loaded-language propaganda. In the released dataset, 30% of fake articles use appeal to authority, 30% use loaded language, and 40% contain disinformation without either technique.

  2. Knowl 2 — PROPANEWS training improves detection of human-written disinformation

    data/table

    The authors compared detectors trained on PROPANEWS, its unvalidated silver version, and competing generated datasets, then evaluated them on human-written disinformation collected from SNOPES and POLITIFACT. Results are AUC percentages, reported as mean ± standard deviation over four runs. PROPANEWS has the highest AUC in all four detector–test-set combinations. Removing either propaganda technique lowers performance relative to full PROPANEWS, while removing both still leaves performance competitive with or above the other generation approaches. Significance markers indicate paired-bootstrap significance over previous best approaches: * p<.05p < .05 and ** p<.01p < .01.

    Training data PolitiFact RoBERTa-Large PolitiFact Grover-Large SNOPES RoBERTa-Large SNOPES Grover-Large
    Without human validation (silver)
    GROVER-GEN 57.65 (±7.6) 52.77 (±2.1) 48.42 (±2.2) 49.53 (±0.1)
    GROVER-GEN-1SENT 49.65 (±5.2) 47.48 (±1.8) 44.44 (±3.2) 50.10 (±2.1)
    FAKEEVENT 46.33 (±2.6) 50.27 (±5.9) 45.36 (±1.2) 47.40 (±1.3)
    FAKEEVENT-1SENT 47.32 (±3.2) 50.12 (±3.2) 46.62 (±2.9) 47.29 (±2.7)
    FACTGEN 48.46 (±2.2) 51.79 (±3.6) 41.98 (±5.4) 50.47 (±4.9)
    FACTGEN-1SENT 41.19 (±3.5) 40.92 (±4.1) 40.01 (±3.8) 45.52 (±3.7)
    PN-SILVER 60.39* (±3.9) 55.23* (±5.8) 51.52** (±3.4) 52.39** (±4.1)
    With human validation (gold)
    PROPANEWS 65.34** (±4.5) 60.43** (±6.2) 53.03** (±3.7) 54.09** (±2.8)
    Without appeal to authority 63.21** (±3.2) 58.28** (±4.2) 50.78* (±1.8) 53.22** (±3.7)
    Without loaded language 64.65** (±1.8) 56.93** (±5.3) 51.92** (±3.4) 51.68* (±1.4)
    Without either technique 61.83* (±4.9) 52.82 (±3.3) 52.77** (±2.7) 50.93 (±2.7)

  3. Knowl 3 — NLI-guided self-critical sequence training discourages entailed replacements

    equation

    For an article xx, let y∗y^* be the original salient sentence to be replaced, and let x~\tilde{x} be the article with y∗y^* masked. BART encodes x~\tilde{x} as heh_e. During training, the masked-out sentence is also used as the maximum-likelihood target, with loss LmL_m. For reinforcement training, y′y' is a replacement sampled by nucleus sampling with p=0.96p=0.96, and y′′y'' is the greedy-decoded baseline. A RoBERTa-large NLI model fine-tuned on MultiNLI estimates Pnli(y∗,y)P_{nli}(y^*,y), the probability that the premise y∗y^* entails a candidate sentence yy. The sampled sequence receives reward r(y′)=−Pnli(y∗,y′)r(y')=-P_{nli}(y^*,y'); the baseline receives r(y′′)=−Pnli(y∗,y′′)r(y'')=-P_{nli}(y^*,y'').

    Lm=−∑t=1Tlog⁡P(yt∗∣y<t∗,he),Ls=−(r(y′)−r(y′′))∑t=1Tlog⁡P(yt′∣y<t′,he),Lfinal=αLm+βLs.L_m=-\sum_{t=1}^{T}\log P(y^*_t\mid y^*_{<t},h_e),\qquad L_s=-(r(y')-r(y''))\sum_{t=1}^{T}\log P(y'_t\mid y'_{<t},h_e),\qquad L_{final}=\alpha L_m+\beta L_s.

    Here tt indexes tokens, TT is the sequence length, and P(⋅)P(\cdot) is BART’s next-token probability. The paper uses α=1\alpha=1 and β=0.01\beta=0.01. The reward favors replacements that the original sentence does not entail; a separate NLI-based post-processing filter removes candidates still predicted to be entailed. With the self-critical loss, the authors report that the invalid-generation rate falls from 7.8% to 3.2%.

  4. Knowl 4 — Appeal-to-authority generation uses contextual and Wikidata experts

    model/method

    To generate an appeal to authority, the method assembles candidate authorities from two sources: Wikidata experts grouped by occupation and person names recognized in the article by a named-entity tagger. For Wikidata candidates, entities born before 1940 are excluded; within each occupation, the 100 entities with the most outgoing Wikidata statements are retained. The fine-tuned BART infiller is prompted to generate a statement attributed to each candidate, using a prefix such as “ziz_i confirmed that ‘”. The candidate sequence with the lowest perplexity is selected.

    To reduce repetitive phrasing, the method varies subject–verb–object order, selects attribution verbs from {said, concluded, confirmed, emphasized, stated, argued}, and can append a preposition from {on, at, in} before prompting BART to generate further context. The paper reports that the structural, verb, and context variations are each applied 50% of the time. The resulting text presents a generated claim as a statement by an authority.

  5. Knowl 5 — Loaded-language generation uses a two-stage infilling procedure

    model/method

    The authors trained a separate BART model to add loaded language, such as emotion-triggering adjectives or adverbs. They began with 2,547 fragment-level loaded-language examples and retained 1,017 in which dependency parsing identified an adverb modifying a verb or an adjective modifying a noun. A direct infilling approach often changed surrounding text: it failed to reproduce the unmasked context exactly about 25% of the time. Their alternative first trains BART to insert a mask at the target location in a specially marked input document, then trains it to infill that mask without the self-critical sequence objective. This two-stage procedure reduced failures to reproduce the unmasked context to around 2%.

  6. Knowl 6 — PROPANEWS pairs authentic articles with validated, localized disinformation

    data/table

    PROPANEWS contains 2,256 distinct articles, balanced between authentic articles and manipulated fake articles. Its source is TIMELINE17: 4,535 articles from trustworthy outlets organized into 17 news-event timelines. The source was chosen to retain reliable surrounding content and cover consequential events. In each fake article, the generated disinformation replaces a salient sentence; the dataset records the location of the disinformative passage. Fake articles are divided into three groups: 30% with appeal to authority, 30% with loaded language, and 40% with neither technique. The train, validation, and test split sizes are 1,256, 500, and 500 articles, respectively.

  7. Knowl 7 — Crowd validation creates a gold subset and yields moderate-to-high agreement

    experimental setup

    Amazon Mechanical Turk workers checked generated passages for inaccuracy by looking for supporting evidence in trustworthy news sources. Only passages judged inaccurate were retained in the human-validated PROPANEWS data; accurate generated cases were discarded. Around 400 unique workers completed approximately 2,000 annotation tasks. The authors measured agreement using Worker Agreement With Aggregate (WAWA), comparing each worker’s response with the majority-vote aggregate and micro-averaging across examples. WAWA precision was 80.01%, recall was 78.94%, and F1 was 79.47%, which the paper characterizes as moderate-to-high agreement.

  8. Knowl 8 — Human ratings find a small plausibility advantage for PROPANEWS generations

    empirical result

    For a human assessment of plausibility, three Mechanical Turk workers rated each of 100 PROPANEWS-generated articles and each of 100 GROVER-GEN articles on a three-point scale. The average score was 2.25 for PROPANEWS and 2.15 for GROVER-GEN, a slight advantage for PROPANEWS. Among PROPANEWS articles rated highly plausible, 29.2% of workers said the generated propaganda strongly affected their plausibility judgment. These ratings support the authors’ claim that the propaganda additions can contribute to the perceived plausibility of generated articles.

  9. Knowl 9 — Detector errors reflect gaps in static, dynamic, and cross-document knowledge

    empirical result

    In an analysis of human-written disinformation that detectors misclassified, the authors attributed about 30% of errors to missing static knowledge available in public resources, about 48% to difficulty acquiring dynamic information from newer news sources, and the remainder to missing multi-document reasoning. As one indication of the temporal problem, RoBERTa achieved 69.0% accuracy on fake articles published before 2019 but 51.9% on articles published after 2019. These are the authors’ approximate error-category estimates and their reported accuracy comparison.

  10. Knowl 10 — Generated data still differs from human-authored disinformation

    limitation

    Increasing the size of PN-SILVER initially improved detector performance, but performance plateaued once the training set was expanded to five times its original size. The authors interpret this as evidence of a remaining gap between generated and human-crafted disinformation, potentially involving writing style, intent, and the fact that their generated content is entirely false whereas human-authored disinformation may mix false claims with accurate material. The generator covers only appeal to authority and loaded language rather than the full range of propaganda techniques, and it produces English-language news only.

Coverage note — Intermediate pretraining on CNN/Daily Mail and the benchmark of detector performance on PROPANEWS are omitted as supporting training and evaluation details; the paper’s dual-use and ethical discussion is also omitted because it does not add a separate method or empirical finding.

References

  1. 1.Marco T Bastos and Dan Mercea. 2019. The Brexit botnet and user-generated hyperpartisan news. Social science computer review, 37(1):38–54.
  2. 2.Taylor Berg-Kirkpatrick, David Burkett, and Dan Klein. 2012. An empirical investigation of statistical significance in NLP. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 995–1005, Jeju Island, Korea. Association for Computational Linguistics.
  3. 3.Antoine Bosselut, Asli Celikyilmaz, Xiaodong He, Jianfeng Gao, Po-Sen Huang, and Yejin Choi. 2018. Discourse-aware neural rewards for coherent text generation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 173–184, New Orleans, Louisiana. Association for Computational Linguistics.
  4. 4.Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46.
  5. 5.Giovanni Da San Martino, Seunghak Yu, Alberto Barrón-Cedeño, Rostislav Petrov, and Preslav Nakov. 2019. Fine-grained analysis of propaganda in news article. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5636–5646, Hong Kong, China. Association for Computational Linguistics.
  6. 6.Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics.
  7. 7.Hernawan Dewatana and Siti Ummu Adillah. 2021. The effectiveness of criminal eradication on hoax information and fake news. Law Development Journal, 3(3):513–520.
  8. 8.Dimitar Dimitrov, Bishr Bin Ali, Shaden Shaar, Firoj Alam, Fabrizio Silvestri, Hamed Firooz, Preslav Nakov, and Giovanni Da San Martino. 2021. Detecting propaganda techniques in memes. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6603–6617, Online. Association for Computational Linguistics.
  9. 9.Chris Donahue, Mina Lee, and Percy Liang. 2020. Enabling language models to fill in the blanks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2492–2501, Online. Association for Computational Linguistics.
  10. 10.Zhida Feng, Jiji Tang, Jiaxiang Liu, Weichong Yin, Shikun Feng, Yu Sun, and Li Chen. 2021. Alpha at SemEval-2021 task 6: Transformer based propaganda classification. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 99–104, Online. Association for Computational Linguistics.
  11. 11.Yi Fung, Christopher Thomas, Revanth Gangi Reddy, Sandeep Polisetty, Heng Ji, Shih-Fu Chang, Kathleen McKeown, Mohit Bansal, and Avi Sil. 2021. InfoSurgeon: Cross-media fine-grained information consistency checking for fake news detection. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1683–1698, Online. Association for Computational Linguistics.
  12. 12.Karl Moritz Hermann, Tomás Kociský, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, pages 1693–1701, Montreal, Quebec, Canada.
  13. 13.Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. In Proceedings of the 8th International Conference on Learning Representations, ICLR ’20, Addis Ababa, Ethiopia. OpenReview.net.
  14. 14.Hamid Karimi and Jiliang Tang. 2019. Learning hierarchical discourse-level structure for fake news detection. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3432–3442, Minneapolis, Minnesota. Association for Computational Linguistics.
  15. 15.Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations, ICLR ’15, San Diego, CA, USA.
  16. 16.Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics.
  17. 17.Yang Liu and Mirella Lapata. 2019. Text summarization with pretrained encoders. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3730–3740, Hong Kong, China. Association for Computational Linguistics.
  18. 18.Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv:1907.11692.
  19. 19.Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In Proceedings of the 7th International Conference on Learning Representations, ICLR ’19, New Orleans, LA, USA. OpenReview.net.
  20. 20.Giovanni Da San Martino, Stefano Cresci, Alberto Barrón-Cedeño, Seunghak Yu, Roberto Di Pietro, and Preslav Nakov. 2020. A survey on computational propaganda detection. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 4826–4832. ijcai.org.
  21. 21.Pedro Henrique Martins, Zita Marinho, and André F. T. Martins. 2020. Sparse text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4252–4273, Online. Association for Computational Linguistics.
  22. 22.Terufumi Morishita, Gaku Morio, Hiroaki Ozaki, and Toshinori Miyoshi. 2020. Hitachi at SemEval-2020 task 3: Exploring the representation spaces of transformers for human sense word similarity. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 286–291, Barcelona (online). International Committee for Computational Linguistics.
  23. 23.Ahmadreza Mosallanezhad, Kai Shu, and Huan Liu. 2021. Generating topic-preserving synthetic news. In Proceedings of the 2021 IEEE International Conference on Big Data, Big Data ’21, pages 490–499.
  24. 24.Van-Hoang Nguyen, Kazunari Sugiyama, Preslav Nakov, and Min-Yen Kan. 2020. FANG: leveraging social context for fake news detection using graph representation. In Proceedings of the 29th ACM International Conference on Information and Knowledge Management, CIKM ’20, pages 1165–1174, Ireland (Virtual Event). ACM.
  25. 25.Qiang Ning, Hao Wu, Rujun Han, Nanyun Peng, Matt Gardner, and Dan Roth. 2020. TORQUE: A reading comprehension dataset of temporal ordering questions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1158–1172, Online. Association for Computational Linguistics.
  26. 26.Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1999. The PageRank citation ranking: Bringing order to the web. Technical Report 1999-66, Stanford InfoLab.
  27. 27.Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers, John Thickstun, Sean Welleck, Yejin Choi, and Zaïd Harchaoui. 2021. MAUVE: measuring the gap between neural text and human text using divergence frontiers. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS ’21, pages 4816–4828.
  28. 28.Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.
  29. 29.Hannah Rashkin, Eunsol Choi, Jin Yea Jang, Svitlana Volkova, and Yejin Choi. 2017. Truth of varying shades: Analyzing language in fake news and political fact-checking. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2931–2937, Copenhagen, Denmark. Association for Computational Linguistics.
  30. 30.Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. 2017. Self-critical sequence training for image captioning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR ’17, pages 1179–1195, Honolulu, HI, USA. IEEE Computer Society.
  31. 31.Niloufar Salehi, Lilly Irani, Michael S. Bernstein, Ali Alkhatib, Eva Ogbe, Kristy Milland, and Clickhappier. 2015. We are dynamo: Overcoming stalling and friction in collective action for crowd workers. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, CHI ’15, pages 1621–1630, Seoul, Republic of Korea. ACM.
  32. 32.Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2021. “nice try, kiddo”: Investigating ad hominems in dialogue responses. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 750–767, Online. Association for Computational Linguistics.
  33. 33.Kai Shu, Yichuan Li, Kaize Ding, and Huan Liu. 2021. Fact-enhanced synthetic news generation. In Proceedings of the Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI ’21’, pages 13825–13833. AAAI Press.
  34. 34.Kai Shu, Deepak Mahudeswaran, Suhang Wang, Dongwon Lee, and Huan Liu. 2018. FakeNewsNet: A data repository with news content, social context and dynamic information for studying fake news on social media. ArXiv:1809.01286.
  35. 35.Giang Binh Tran, Tuan Tran, Nam Khanh Tran, Mohammad Alrifai, and Nattiya Kanhabua. 2013. Leveraging learning to rank in an optimization framework for timeline summarization. In Proceedings of the SIGIR 2013 Workshop on Time-aware Information Access, TAIA ’13.
  36. 36.Sander van Der Linden, Jon Roozenbeek, and Josh Compton. 2020. Inoculating against fake news about COVID-19. Frontiers in psychology, 11:2928.
  37. 37.Herman Wasserman and Dani Madrid-Morales. 2019. An exploratory study of “fake news” and media trust in Kenya, Nigeria and South Africa. African Journalism Studies, 40(1):107–123.
  38. 38.Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112–1122, New Orleans, Louisiana. Association for Computational Linguistics.
  39. 39.Ronald J Williams. 1992. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229–256.
  40. 40.Xueqing Wu, Kung-Hsiang Huang, Yi Fung, and Heng Ji. 2022. Cross-document misinformation detection based on event graph reasoning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 543–558, Seattle, United States. Association for Computational Linguistics.
  41. 41.Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, Franziska Roesner, and Yejin Choi. 2019. Defending against neural fake news. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS ’19, pages 9051–9062, Vancouver, BC, Canada.

Citation

MLA
Huang, K.-H., et al. “Faking Fake News for Real Fake News Detection: Propaganda-Loaded Training Data Generation”. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2023, pp. 14571–89, https://doi.org/10.18653/v1/2023.acl-long.815.
APA
Huang, K.-H., McKeown, K., Nakov, P., Choi, Y., & Ji, H. (2023). Faking Fake News for Real Fake News Detection: Propaganda-Loaded Training Data Generation. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 14571–14589. https://doi.org/10.18653/v1/2023.acl-long.815
Chicago
Huang, K.-H., K. McKeown, P. Nakov, Y. Choi, and H. Ji. 2023. “Faking Fake News for Real Fake News Detection: Propaganda-Loaded Training Data Generation”. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 14571–89. https://doi.org/10.18653/v1/2023.acl-long.815.
Harvard
Huang, K.-H. et al. (2023) “Faking Fake News for Real Fake News Detection: Propaganda-Loaded Training Data Generation”, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp. 14571–14589. Available at: https://doi.org/10.18653/v1/2023.acl-long.815.
Vancouver
1. Huang K-H, McKeown K, Nakov P, Choi Y, Ji H (2023) Faking Fake News for Real Fake News Detection: Propaganda-Loaded Training Data Generation. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp 14571–14589

BibTeX

@inproceedings{huang-etal-2023-faking,
    title = "Faking Fake News for Real Fake News Detection: Propaganda-Loaded Training Data Generation",
    author = "Huang, Kung-Hsiang  and
      McKeown, Kathleen  and
      Nakov, Preslav  and
      Choi, Yejin  and
      Ji, Heng",
    editor = "Rogers, Anna  and
      Boyd-Graber, Jordan  and
      Okazaki, Naoaki",
    booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.acl-long.815/",
    doi = "10.18653/v1/2023.acl-long.815",
    pages = "14571--14589"
}
Metadata:ACL Anthology

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: https://creativecommons.org/licenses/by/4.0/