Multi-label mental health classification in social media posts with multi-perspective prompt ensemble and auxiliary self-supervision

Abstract

Anxiety and depression have become major global health concerns. With the rapid rise of social media, people increasingly share emotions and personal struggles through posts, which often convey multiple mental states simultaneously. To address this multi-label classification challenge in mental health texts, this study proposes a multi-task framework with two main modules, a multi-perspective prompt design module and a perturbation-based self-supervised learning module, based on a pre-trained language model backbone. Prompts from sociological, psychological, and educational perspectives are used to enhance semantic understanding. To improve model robustness, we formulate self-supervised auxiliary tasks where the model predicts whether a sentence has undergone insertion, swap, or deletion. Experiments on the MultiWD dataset, covering six wellness dimensions, show that our method outperforms all baselines. Furthermore, ablation studies explore the impact of different training configurations and confirm the critical contributions of both proposed modules.