Authorship Profiling in Social Media – From a Mental Health Perspective

Dr. Jonathan Schler, Chair of Computer Science Department, HIT

Session I: March 30^th, 09:40-10:00

Attempts to predict mental health vulnerabilities in everyday social media messages have proliferated rapidly over the last decade. In this talk we present methods of profiling mental health status based on language used in a neutral setting (i.e., where people are not discussing mental health), we used an array of machine learning and regression models to classify posts from neutral Reddit forums as written by depressed users (self-identified) or random controls. Predictive linguistic features included dictionary-based categories (e.g., LIWC variables), parts of speech, and character n-grams. The best-performing classifier was a BERT based model, which is not interpretable in a traditional sense by practitioners or mental health technology users. We discuss the possibilities of reverse-engineering psychological insights from “black box” machine learning models and otherwise using these models to advance theory in behavioral science. Finally, we consider ethical and practical considerations of our findings for mental health technology applications.