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

DualDet Dataset Paper DualDet: A Dual-Task Detection Benchmark for Stance and Bot Detection on Social Media. Introduction DualDet integrates inherited bot labels, follower graph context, and new expert annotations for user-level stance. The published dataset covers 124,802 users, with 22,906 users carrying expert stance labels. The benchmark supports the experiments reported in the paper, including stance detection, bot detection, and joint modeling, and highlights measurable stance–bot dependencies. Dataset Format DualDet Location: DualDet.zip/DualDet/unlabeled Contents: each entry corresponds to a single tweet and records author_id, id, and text exactly as extracted from TwiBot-22. Note: the id column here refers to the tweet id (per-tweet data). labeled subset Location: DualDet.zip/DualDet/labeled Contents: each record summarizes one user with hashed id, hashed name, tweets (embeddings as generated), follow (hashed follower ids), and stance. This view supports user-centric sentiment or stance modelling. Note: the id field in this section refers to the user id after hashing. test Location: DualDet.zip/DualDet/test Contents: rows combine id, name, description, tweets, follow, stance, and account-type. Note: in experiments, we conduct evaluation using a 70/15/15 train/validation/test split with a strict user-level partition, description and tweets store BERT embeddings, while every identifier (id, name, follow) remains hashed through the shared HMAC. How to obtain The processed data above is already available in this workspace. For access to more detailed or raw data releases, please contact us at 2510263007@mails.szu.edu.cn using your institutional email. In your request, include your institution, advisor (if any), and intended research use. Ethics & Anonymization DualDet is built from publicly available social media content and licensed under CC BY 4.0 for research use. The corpus may contain sensitive or offensive language, and exposing user-level information increases privacy risks. Here we publish identifier-based outputs only: id, name, and follow are hashed with a keyed HMAC, while description/tweets fields are replaced with BERT embeddings. Please release/use the data in a platform-compliant way, report results at aggregate levels, and audit models for robustness/bias. Misuse risks include political profiling, harassment, or wrongful bot labels—careful handling and ethical review are encouraged before downstream deployment.