Automatically Finding Evidence and Predicting Answers in Mental Health Self-Report Questionnaires

This paper describes the participation of the RELAI team in the eRisk 2024 shared tasks related to the search for symptoms of depression (T1) and the measure of severity of the signs of eating disorders (T3). Both tasks centered on self-report questionnaires and social media textual content. In T1, sentences relevant to each item in a standard depression were mined from a large set of sentences collected from social media. Our approach relied on ground-truth relevance judgments to train multi-label classifiers in deciding whether a sentence is relevant to each item. The goal of T3 was to automatically fill out a standard eating disorder questionnaire based on histories of writings from social media. Given the small set of annotated data and large output space, our approach proceeded by making global, aggregate predictions first and use those predictions to make precise per-item predictions.