| Abstract | Research on automated detection of regrettable disclosures in online social networks (OSN) is limited by the lack of large-scale, semantically rich, and fine-grained annotated datasets. Existing datasets often provide narrow coverage and conflate regret with related phenomena such as toxicity or hate speech, hindering robust modeling of regret-specific cues. To address these limitations, we introduce ReDoS-M, a large-scale, multisource corpus constructed via a hybrid annotation pipeline that combines crowd-sourced labeling, transformerbased self-training, and enrichment with Sentiment-Moral-Emotion (SME) features. Starting from a collection of more than 5.5M user-generated posts and comments gathered from platforms such as Reddit and X (formerly Twitter), we derive four complementary corpora ranging from 4.27M to 5.13M annotated items, reflecting different annotation and label-fusion strategies. We evaluate ReDoS-M in terms of label coverage and downstream utility by training and evaluating six transformer-based models (DeBERTa and XLM-RoBERTa variants, with and without SME and Large Language Model-generated features). Across the ReDoS-M corpora, all six models achieve strong performance, with micro-F1 scores exceeding 0.98 and AUC values above 0.99 in the best settings, demonstrating that ReDoS-M supports effective and generalizable detection of regrettable OSN disclosures. Overall, ReDoS-M constitutes a comprehensive and scalable foundation for advancing research on fine-grained modeling and classification of regrettable disclosures in OSN environments. |
|---|
