Do text-based hooks predict engagement? A measurement framework and a cautionary null result for ICH short videos on Douyin

Intangible Cultural Heritage (ICH) promotion on social media faces unique challenges, especially on feed-based, short- video platforms like Douyin. This study examines Intangible Cultural Heritage (ICH) videos on Douyin by decomposing the concept of a “hook strategy” into 24 distinct features and ranking their relative importance to user engagement. Our contribution is primarily a measurement framework that decomposes the otherwise loosely defined notion of a “hook” into concrete, reproducible textual and verbal features, together with a comparison of how strongly these features are associated with engagement in a large ICH corpus. We do not aim to predict engagement—which in this cultural-communication setting is inherently high-variance—but to compare the relative importance of these features. The findings reveal that importance concentrates in title-level and title–text features—Title Sentiment Score (TSS), Title–Text Semantic Similarity (TtSS), and a mutually-collinear semantic-magnitude block (Title and Text Semantic Density and their difference, TSD, tSD and TtSD). Because the random forest cannot separate this collinear block and does not generalize out of sample (held-out R 2 ≈ 0), we test these associations in a covariate-adjusted regression layer that controls for length and corrects for multiple testing. Only Title Semantic Density survives: concise, focused titles (higher title-embedding magnitude) are associated with higher likes and comments after adjusting for length ( β ≈ 0.30–0.35, FDR- p < 0.02). The remaining high-importance features—including title sentiment—are not robust to adjustment, and unadjusted correlations are weak because several measures co-vary with length. We read this gap between importance rank and adjusted association as a substantive result: most text-based hook signals do not survive proper adjustment, and engagement is dominated by exposure. The contribution is thus a reproducible measurement framework, one robust association (title focus), a construct-validity critique of norm-based semantic measures, and a cautionary negative result for text-only accounts of short-video engagement, pending experimental validation.