Mental Health Prediction from Social Media Text Using Mixture of Experts

Predicting mental health statuses from social media text is a well-known Natural Language Processing (NLP) task. In this work, we focus on the issue of depression and anxiety disorder prediction from Twitter by comparing a more conventional approach based on engineered features with a data-oriented alternative based on mixture of specialists with transformer language models. Results from a large corpus of depression/anxiety self-disclosed diagnoses in the Portuguese language are reported, and a feature importance analysis is carried out to provide further insights into these tasks.