From Words to Actions: Natural Language Processing for suicide prevention helplines

This dissertation explores the integration of NLP and ML techniques to enhance the effectiveness and efficiency of suicide prevention helplines. By developing and evaluating AI-driven support tools, this research aims to provide real-time assistance to counselors, improve the quality of helpline services, and gain deeper insights into the challenges faced by individuals in crisis. Chapter 2 outlines the design and development of a content-based recommender support system aimed at assisting counselors in a suicide prevention helpline. The primary motivation was to address the cognitive and emotional challenges faced by counselors during intense chat conversations. By implementing a system that provides suggestions based on previous successful interactions, the goal was to investigate the potential for improved counselor efficacy. Similar interactions were found using averaged word vectors. The evaluation of this tool in a simulated environment showed that the usability of the tool was good and the retrieved suggestions were better than random suggestions, but tailored advice from a human expert still had significantly more utility. Building on the support tool developed in Chapter 2, Chapter 3 presents an evaluation of the effectiveness through a randomized controlled trial. The recommender system was improved using a deep-learning method and tested in real-time counseling sessions to assess its impact on counselors’ self-efficacy and the quality of their responses. While the tool did not significantly improve self-efficacy scores, it was frequently used in longer, more complex conversations, suggesting its utility in challenging situations. The study highlighted the feasibility of integrating AI-assisted tools in helpline services, paving the way for future enhancements. Chapter 4 focuses on evaluating various topic modeling methods to analyze conversation data from mental health helplines. Traditional methods like LDA were compared with newer techniques such as BERTopic, which leverages sentence embeddings for better context capture. The study found that BERTopic outperformed other methods in terms of topic coherence and interpretability, especially when applied to short, context-dependent texts typical of helpline conversations. The insights gained from this analysis can help helplines better understand the issues faced by their clients and improve their services. The aim of Chapter 5 is to identify changes in conversation topics on a suicide prevention helpline during the COVID-19 pandemic. Using BERTopic, the study analyzed chat data before and after the lockdown measures were implemented. The results indicated significant shifts in conversation topics, with an increased mention expressions of gratitude towards counselors. There were decreases in mentions of specific suicide plans, however, specifically helpseekers who lived alone showed an increase in plans for suicide. These findings underscore the impact of the pandemic on mental health and the potential for monitoring of helpline conversations. Chapter 6 explores the use of AI models to classify counselor and client behaviors in the context of Motivational Interviewing (MI) during helpline chats. By training models on a coded dataset of MI sessions, the study aimed to automate the identification of counseling techniques. The deep learning model BERTje showed high accuracy in classifying MI behaviors, indicating its potential as a tool for providing feedback to counselors. This automation could enhance the monitoring of MI in online helplines and potentially boost adherence. Finally, Chapter 7 investigates which counselor utterances contribute to positive outcomes for help seekers using deep learning models. By analyzing chat logs and help seeker self-assessments, the study identified key behaviors of counselors that positively or negatively affected client wellbeing. Positive affirmations and expressing involvement were linked to improved scores, while the use of macros and premature conversation endings had negative effects. The study highlighted the potential of ML to provide actionable insights for training counselors and enhancing the effectiveness of helpline conversations.