Misinformation Span Detection in Videos via Audio Transcripts
New models pinpoint exactly where false claims appear in videos using audio text.
A team led by Breno Matos and including Savvas Zannettou has tackled a key limitation in video misinformation detection: instead of just classifying a whole video as true or false, their method identifies the exact segments where false claims appear. They built two new datasets from over 500 videos, transcribing audio to text and annotating more than 2,400 segments with fact-checked claims. Using state-of-the-art language models, they achieved an F1 score of 0.68 in detecting these misinformation spans, a significant step toward more precise and explainable fact-checking.
This approach addresses the growing challenge of video-based misinformation on platforms like YouTube and TikTok, where false narratives can be embedded in long-form content. By releasing their datasets, transcripts, audio, and videos publicly, the researchers enable further work on this task. The work was accepted at ICWSM 2026 and represents a move from coarse video-level classification to fine-grained, span-level detection that could help fact-checkers and platforms flag misleading content more effectively.
- Two new datasets with 500+ videos and 2,400+ annotated segments for misinformation span detection
- State-of-the-art language models achieve 0.68 F1 score in identifying false claim segments
- Public release of all datasets, transcripts, audio, and videos to enable further research
Why It Matters
Enables precise, interpretable fact-checking of video misinformation, moving beyond binary classification to pinpoint false claims.