‹ back to 2026-07-16

TikStance: A Multimodal and Hierarchical Dataset for Multi-target Stance Analysis in TikTok Political Conversations

arXiv cs.CL Multimodal & Generative Yazhi Zhang, Fuqiang Niu, Bowen Zhang 2026-07-16

TL;DR - TikStance is a new multimodal, context-aware TikTok dataset for multi-target political stance detection, pairing 161 host videos with hierarchical comment trees to enable analysis grounded in both audiovisual and conversational context.

Contains 161 videos and 13,876 comments centered on three 2024 U.S. election figures (Trump, Biden, Harris), collected September 2023–January 2025.
Each discussion unit links a host video plus metadata to a parent-linked comment tree, supporting video-to-target and comment-to-target stance analysis.
Labeled by three annotators on a three-class scheme (Favor, Against, None), with disagreements re-annotated; Krippendorff's α reached 0.743 (Trump), 0.723 (Biden), 0.722 (Harris).
Descriptive analysis shows target-dependent stance distributions and conversational depth, with nested replies making up 23.3% of comments.

view merged work →