BERT-Based Transformers for Early Detection of Mental Health Illnesses

Abstract

This paper briefly describes our research groups' efforts in tackling Task 1 (Early Detection of Signs of Self-Harm), and Task 2 (Measuring the Severity of the Signs of Depression) from the CLEF eRisk Track. Core to how we approached these problems was the use of BERT-based classifiers which were trained specifically for each task. Our results on both tasks indicate that this approach delivers high performance across a series of measures, particularly for Task 1, where our submissions obtained the best performance for precision, F1, latency-weighted F1 and ERDE at 5 and 50. This work suggests that BERT-based classifiers, when trained appropriately, can accurately infer which social media users are at risk of self-harming, with precision up to 91.3% for Task 1. Given these promising results, it will be interesting to further refine the training regime, classifier and early detection scoring mechanism, as well as apply the same approach to other related tasks (e.g., anorexia, depression, suicide).