Dataset Generation Framework Guided by the Online Gambling Disorder Questionnaire

Gambling disorder is a public health concern, but research on the early detection of gambling-related harm is limited by the scarcity of ethically shareable online conversations. This study presents a framework for generating an entirely synthetic dataset for natural language processing research on gambling-related behavioral signals. Behavioral dimensions from the Online Gambling Disorder Questionnaire were used to construct 2100 risk-aligned user profiles and generate X/Twitter-style monologues and multi-user threads across gambling and non-gambling subtopics. The dataset contains 35,346 monologues and 1716 threads. The generated text was evaluated using automatic metrics, manual target-consistency assessment, and downstream classification under user-disjoint and parent-topic-disjoint protocols. Automatic and manual evaluations indicated acceptable linguistic quality and consistency with predefined behavioral targets. In the stricter parent-topic-disjoint setting, logistic regression achieved a macro-F1 of 0.884 for binary gambling-content detection, whereas four-level risk prediction remained more difficult. These results show that the dataset contains learnable signals for computational detection of gambling-related content without exposing real-user data. However, the dataset was not externally validated against authentic social-media conversations and is not intended for clinical diagnosis. The resulting resource is designed to support reproducible research on gambling disorder and mental health while preserving user privacy.