Mental Health Analysis Through Social Media Using a Large Language Model

Abstract

Mental health is a global issue. Mental health in individuals has been continuously deteriorating over the past few years. Its early diagnosis and treatment can help improve individual physical and emotional health. Large language models (LLMs) are influencing various fields, including medicine. The recent advent of LLMs offers much hope for cost- and time-efficient scaling of monitoring and intervention. These advancements in LLMs have empowered various applications. However, there is a significant gap in research regarding analyzing and enhancing the potentialities of LLMs in the mental health domain. Emergent abilities were recently discovered as an outcome of scaling up language models. They span a variety of language models, task types, and experimental scenarios. These emergent abilities seem to be nontrivial upcoming research directions for the domain of Natural Language Processing. This study analyzes these abilities, for which meaningful performance has only been observed at a certain computational scale. In this work, I comprehensively evaluated LLaMA-2 on various mental health detection tasks using various prompt engineering methods, zero-shot, and few-shot prompting. To enhance LLaMA-2’s capability for mental health analysis, the model is further fine-tuned to the social media mental health dataset by applying the parameter-efficient fine-tuning (PEFT) methods LoRA and QLORA. The study’s main objectives include fine-tuning and evaluating the LLaMa-2 for mental disorder detection, comparing its performance with the emergent capabilities of LLM, prompt engineering, and examining factors that contribute to its success. The model’s robustness and generalization capabilities are also tested on a publicly available mental health dataset. The results indicate a promising yet limited performance of LLaMA-2-7b with zero-shot and few-shot prompt designs for mental health tasks. More importantly, the experiments show that instruction fine-tuning can significantly improve the performance of LLMs for all tasks simultaneously. Furthermore, this research contributes to various potential applications in different domains, including mental illness analysis for early detection and efficient treatment by practitioners, personalized recommendations for self-management of mental health conditions and trends, and risk factor tracking associated with men- tal health disorders at the population level for public health planning and policymaking. The thesis concludes with recommendations for future work.