UPV-Symanto at eRisk 2021: Mental Health Author Profiling for Early Risk Prediction on the Internet

Abstract

[EN] This paper presents the contributions of the UPV-Symanto team, a collaboration between Symanto Research and the PRHLT Center, in the eRisk 2021 shared tasks on gambling addiction, self-harm detection and prediction of depression levels. We have used a variety of models and techniques, including Transformers, hierarchical attention networks with multiple linguistic features, a dedicated early alert decision mechanism, and temporal modelling of emotions. We trained the models using additional training data that we collected and annotated thanks to expert psychologists. Our emotions-over-time model obtained the best results for the depression severity task in terms of ACR (and second best according to ADODL). For the self-harm detection task, our Transformer-based model obtained the best absolute result in terms of ERDE5 and we ranked equal first in terms of speed and latency.