1,827 labeled short videos from Douyin categorized as real or fake news for multimodal detection tasks. The dataset provides visual frames, audio signals, ASR transcripts, OCR text, and social context including user metadata and comments.
Use Cases
- Train multimodal fake news detectors using the ASR and OCR text features combined with visual frame embeddings
- Evaluate the influence of social context on detection accuracy using the publisher metadata and user comment columns
- Perform cross-modal alignment studies between audio tracks and visual content to detect manipulated media
Strengths
- 1,827 video samples with ground-truth labels verified by professional fact-checking organizations
- Multimodal features including ASR-derived text, OCR-extracted text, and raw audio/visual streams
- Social context data encompassing publisher information and hierarchical user comment structures