Detection of Mental Health from Reddit via Deep Contextualized Representations

We address the problem of automatic detection of psychiatric disorders from the linguistic content of social media posts. We build a large scale dataset of Reddit posts from users with eight disorders and a control user group. We extract and analyze linguistic characteristics of posts and identify differences between diagnostic groups. We build strong classification models based on deep contextualized word representations and show that they outperform previously applied statistical models with simple linguistic features by large margins. We compare user-level and post-level classification performance, as well as an ensembled multiclass model.