Full paper
ShortCheck: Checkworthiness Detection of Multilingual Short‐Form Videos
By Henrik Vatndal, Vinay Setty
The problem
Short-form videos such as TikTok, YouTube Shorts, and Instagram Reels have transformed how information (including mis- and disinformation) spreads. Importantly, the chaotic way of short form videos mixing speech, text overlays, and pictures and videos makes it particularly difficult to spot misinformation in them. What is worse, manual monitoring cannot keep pace with billions of daily posts.
The solution
Factiverse researchers developed ShortCheck, an AI system designed to automatically identify "checkworthy" videos that require human fact-checking. ShortCheck analyzes videos across multiple media formats simultaneously.
It is easy to use and can predict checkworthiness in over 30 major languages. It extracts on-screen text, transcribes spoken audio, summarizes visual scenes, and scans for political buzzwords or synthetic "deepfake" media.
So, Shortcheck integrates speech transcription, Optical Character Recognition (OCR), object and deepfake detection, video-to-text summarisation, and claim verification.
The results
When tested on real-world TikTok datasets, ShortCheck achieved accuracy of over 70% on key performance benchmarks across English and Norwegian content. Interestingly, transcript features and ideological language were the strongest signals of checkworthiness, while visual cues were less important. More importantly, reducing the amount of content generated each day down to a manageable selection cuts both processing time and costs down.
Summing up, with clear, step-by-step explanations for its decisions, ShortCheck can help fact-checkers to prioritize claims that need to be checked quicker, hence make their and fight against viral misinformation more effective.
.avif)
