DualDet: A Dual-Task Detection Benchmark for Stance and Bot Detection on Social Media

Abstract

Social media manipulation is often driven by automated accounts that amplify polarized viewpoints, posing persistent challenges to online information integrity. Two core capabilities for understanding and mitigating such manipulation are stance detection and bot detection. Although usually studied separately, they are coupled in practice: automation can skew stance distributions, while stance-aligned communities can shape visibility and network signals used for bot detection. We introduce DualDet, a dual-task benchmark for stance and bot detection spanning election discourse (Biden, Trump) and vaccine-related discourse (Vaccine). DualDet integrates inherited bot labels and follower-graph context with expert-annotated user-level stance labels, achieving substantial inter-annotator agreement (Cohen’s k > 0.9). The dataset contains 124,802 users in total, of which 22,906 users are annotated with expert stance labels. We further provide dataset analyses and reproducible baselines for stance detection, bot detection, and joint evaluation, revealing measurable stance–bot dependencies and highlighting open challenges for coupling-aware modeling.