Emotion-Infused Models for Explainable Psychological Stress Detection

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2895–2909 June 6–11, 2021. ©2021 Association for Computational Linguistics 2895 Emotion-Infused Models for Explainable Psychological Stress Detection Elsbeth Turcan1 and Smaranda Muresan1,2 and Kathleen McKeown1 Department of Computer Science, Columbia University1 Data Science Institute, Columbia University2 {eturcan, smara, kathy}@cs.columbia.edu Abstract The problem of detecting psychological stress in online posts, and more broadly, of detecting people in distress or in need of help, is a sen- sitive application for which the ability to inter- pret models is vital. Here, we present work exploring the use of a semantically related task, emotion detection, for equally compe- tent but more explainable and human-like psy- chological stress detection as compared to a black-box model. In particular, we explore the use of multi-task learning as well as emotion- based language model fine-tuning. With our emotion-infused models, we see comparable results to state-of-the-art BERT. Our analysis of the words used for prediction show that our emotion-infused models mirror psychological components of stress. 1 Introduction As crises have begun to multiply worldwide, in- cluding the COVID-19 pandemic and the resulting economic downturn, psychological stress has risen dramatically1. The problem of detecting psycholog- ical stress, and more broadly, of detecting people in distress and in need of help, is a sensitive appli- cation; therefore, the ability to interpret the results, in order to understand why, is vital. The conse- quences of blindly trusting a black-box model and mislabeling users’ stress levels could be serious in a deployed application such as a therapeutic chatbot, where some users may not receive the immediate help they need. Furthermore, models that make de- cisions based on psychology theory about factors that impact stress will be easier for humans to un- derstand, and their mistakes will be more obvious. Researchers have recently begun to study psycho- logical stress, but in this work, we propose a new focus on examining the information our models use to make decisions and finding ways to incorporate psychological factors, like emotion, into them. 1https://www.apa.org/news/press/ releases/stress/2020/report-october To approach the problem of stress detection, which has much less labeled data than many pop- ular classification tasks, we first note that stress has been shown to interact with emotion (Lazarus, 2006; Thoern et al., 2016; Levenson, 2019), a task that has far more publicly available labeled data. For example, individuals who are stressed are likely to express emotions such as fear, sadness, or anger and unlikely to express emotions such as happiness. Traditional multi-task learning would normally be helpful in this situation, but there are no cur- rently available datasets labeled with both stress and emotion. Even if there were, it would be bene- ficial to incorporate external information without re-labeling new datasets for each new combination of useful tasks. Here, we present work exploring how to use semantically related tasks–here, emo- tion detection–to create emotion-infused models capable of equally competent, but explainable, psy- chological stress detection as compared to a …

Content truncated to 500 words