Claim Verification with Tiered Trusted Sources: Why Style-Based Fake News Classifiers Fail to Generalize

Author

Ananya Singh Baghel, Tanish Kumar Sahu, Shweta Dubey


Abstract

 Many fake news detectors are text classifiers trained on labelled news articles. They score very high on their own test data, but we suspected that they learn how a publisher writes and not whether a statement is true. This paper tests that idea on LIAR, a set of short political claims, and compares the classifier with a claim-checking pipeline that uses web evidence.

A Random Forest trained on the ISOT dataset reaches 99.76% on ISOT, yet it scores only 44.5% on 200 LIAR claims and calls every one of them fake. The pipeline extracts claims with a large language model (LLM), searches the web with three query variants, keeps only pages from a list of trusted domains, and gives a verdict only when two independent trusted domains agree. We compare it with a plain LLM and with an LLM that uses raw web search.

The plain LLM was the most accurate system (70.5%). The pipeline reached 65.4% on the 52 claims it answered (26.0% coverage), while the plain LLM scored 78.8% on those same 52 claims. The extra query variants raised coverage from 17.0% to 26.0%, but accuracy did not change in a measurable way. We conclude that style classifiers do not carry over to new data, and that a strict source rule gives checkable evidence and caution, not better accuracy than a strong LLM.


Keywords

misinformation detection, claim verification, evidence retrieval, source credibility, ISOT, LIAR



Full Text:

Download Paper PDF


References


  1. Vlachos and S. Riedel, “Fact checking: Task definition and dataset construction,” in ACL Workshop on Language Technologies and Computational Social Science, 2014.
  2. Shu, A. Sliva, S. Wang, J. Tang and H. Liu, “Fake news detection on social media: A data mining perspective,” ACM SIGKDD Explorations Newsletter, vol. 19, no. 1, pp. 22–36, 2017.
  3. Ahmed, I. Traore and S. Saad, “Detection of online fake news using N-gram analysis and machine learning techniques,” in Intelligent, Secure, and Dependable Systems in Distributed and Cloud Environments (ISDDC), Springer, 2017.
  4. Thorne, A. Vlachos, C. Christodoulopoulos and A. Mittal, “FEVER: a large-scale dataset for fact extraction and VERification,” in Proc. NAACL-HLT, 2018.
  5. Schlichtkrull, Z. Guo and A. Vlachos, “AVeriTeC: A dataset for real-world claim verification with evidence from the web,” in NeurIPS Datasets and Benchmarks Track, 2023.
  6. Guo, M. Schlichtkrull and A. Vlachos, “A survey on automated fact-checking,” Transactions of the Association for Computational Linguistics, vol. 10, pp. 178–206, 2022.
  7. Y. Wang, “‘Liar, liar pants on fire’: A new benchmark dataset for fake news detection,” in Proc. 55th Annual Meeting of the ACL (Short Papers), 2017.
  8. Shu, D. Mahudeswaran, S. Wang, D. Lee and H. Liu, “FakeNewsNet: A data repository with news content, social context, and spatiotemporal information for studying fake news on social media,” Big Data, vol. 8, no. 3, 2020.
  9. Nakov et al., “Automated fact-checking for assisting human fact-checkers,” in Proc. IJCAI, 2021.
  10. Lewis et al., “Retrieval-augmented generation for knowledge-intensive NLP tasks,” in Advances in Neural Information Processing Systems (NeurIPS), 2020.
  11. B. Wilson, “Probable inference, the law of succession, and statistical inference,” Journal of the American Statistical Association, vol. 22, no. 158, pp. 209–212, 1927.

 

 

Share your valuable work from Social Media Buttons