Leveraging transformer-based language models for psychological distress classification in social media texts

Amjad, AdeenAdeenAmjadDr. AZHAR MuhammadAbdulrahman, RajaRajaAbdulrahmanDewi, Deshinta ArrovaDeshinta ArrovaDewiJamil, AleenaAleenaJamilShabbir, IrfaIrfaShabbir2026-07-202026-07-202025Amjad, A., Azhar, M., Abdulrahman, R., Dewi, D. A., Jamil, A., & Shabbir, I. (2025). Leveraging transformer-based language models for psychological distress classification in social media texts. In IEEE (Ed.). The proceedings of 2025 international conference on emerging research in computational science (ICERCS). 2025 International Conference on Emerging Research in Computational Science (ICERCS), Coimbatore, India (pp. 1-10). IEEE.97983315478999798331547905http://hdl.handle.net/20.500.11861/28279The rise in the prevalence of social media platforms has necessitated the development of automated software to detect early warning signs of psychological distress in user-generated content. Despite the computational model’s potential, the approach remains primarily geared towards single-task classification, thereby overlooking the multifaceted nature of assessing clinical distress. In this paper, the new transformer-based multi-task learning approach is presented to classify psychological distress across three essential domains: severity, primary clinical domain, and urgency. Notably, the new approach integrates clinical feature maps and the temporal analysis component to provide evidence of symptom progression. The model, tested and validated on a carefully analyzed dataset of 15,000 Reddit posts, achieves state-of-the-art performance, with an accuracy of 85.7% and an AUC-ROC of 0.932, thereby setting new benchmarks for contemporary single-task and multi-task models. Notably, the model’s performance in detecting the urgency level remains particularly strong, an essential component of risk assessment.enSocial MediaPsychological DistressLanguage ModelSocial Media TextTransformer-Based Language ModelsClinical CharacteristicsFeature MapsTemporal AnalysisUser-Generated ContentMulti-Task LearningMulti-Task ModelLevel Of UrgencyMental HealthConvolutional Neural NetworkValidation SetClassification TaskF1 ScoreRecurrent Neural NetworkAttention MechanismRepresentation LearningTemporal ModulationPrimary DimensionsCognitive DistortionsMatthews Correlation CoefficientTransformer ArchitectureMacro F1 ScoreEmotional ValenceTransformer ModelIncorrect LabelsTraditional Machine Learning ApproachesLeveraging transformer-based language models for psychological distress classification in social media textsConference Paper10.1109/ICERCS65898.2025.11580797