SMHD-GER: A Large-Scale Benchmark Dataset for Automatic Mental Health Detection from Social Media in German

Findings of the Association for Computational Linguistics: EACL 2023, pages 1526–1541 May 2-6, 2023 ©2023 Association for Computational Linguistics SMHD-GER: A Large-Scale Benchmark Dataset for Automatic Mental Health Detection from Social Media in German Sourabh Zanwar RWTH Aachen University sourabh.zanwar@rwth-aachen.de Daniel Wiechmann University of Amsterdam d.wiechmann@uva.nl Yu Qiao RWTH Aachen University yu.qiao@rwth-aachen.de Elma Kerz RWTH Aachen University elma.kerz@ifaar.rwth-aachen.de Abstract Mental health problems are a challenge to our modern society, and their prevalence is pre- dicted to increase worldwide. Recently, a surge of research has demonstrated the potential of automated detection of mental health condi- tions (MHC) through social media posts, with the ultimate goal of enabling early interven- tion and monitoring population-level health out- comes in real time. Progress in this area of re- search is highly dependent on the availability of high-quality datasets and benchmark corpora. However, the publicly available datasets for understanding and modeling MHC are largely confined to the English language. In this paper, we introduce SMHD-GER (Self-Reported Men- tal Health Diagnoses for German), a large-scale, carefully constructed dataset for MHC detec- tion built on high-precision patterns proposed for English. We provide benchmark models for this dataset to facilitate further research and conduct extensive experiments. These mod- els leverage engineered (psycho-)linguistic fea- tures as well as BERT-German. We also ex- amine nuanced patterns of linguistic markers characteristics of specific MHC. 1 Introduction Mental health is a major challenge in healthcare and in our modern societies at large (Rehm and Shield, 2019; Santomauro et al., 2021). The World Health Organization estimates that 970 million peo- ple worldwide suffer from mental health condi- tions12, with the rate of undiagnosed mental dis- orders estimated to be as high as 45% (La Vonne et al., 2012). 1https://www.who.int/news-room/fact-sheets/ detail/mental-disorders 2‘Mental disorders’ can also be referred to as ‘mental health conditions’. The latter is sometimes used as a broader term encompassing mental disorders, psycho-social disabilities, and mental conditions, include different types of depression, bipolar disorder, schizophrenia, anxiety disorders, chronic stress etc.. In this work, the two terms are used interchangably. The enormous societal impact of mental health conditions (MHC) requires prevention and inter- vention strategies that focus primarily on screening and early detection. The last decade has seen a surge in digital mental health research, an inter- disciplinary line of research that brings together insights from computational linguistics, cognitive psychology and computational social sciences to understand the relationship between patterns of lan- guage use and mental health conditions (D’Alfonso, 2020; Schindler and Domahidi, 2022). Natural language processing, in particular, is increasingly recognized as having transformative potential to support healthcare professionals in the diagnosis and treatment of mental disorders and enable peo- ple to lead healthy lives (see Guntuku et al. 2017; Thieme et al. 2020; Chancellor and De Choudhury 2020; Zhang et al. 2022 for recent overviews of this research). Progress in this area of research is highly depen- dent on the availability of high-quality datasets and benchmark corpora. Social media has emerged as an increasingly vital resource for obtaining such data, as …

Content truncated to 500 words