Abstract
dc:description.abstractWith the rise of large language models (LLMs), prompt-based learning has become a widely studied and essential topic in the field of Natural Language Processing (NLP). Due to their impressive performance in zero-shot and few-shot settings, prompting offers a strong opportunity for low-resource domains such as mental health. Mental health is a prevalent global concern, with increasing numbers of users sharing their experiences on social media. However, data from these platforms is often complex, and manual annotation at scale is both time-consuming and costly.<br/><br/>This thesis investigates the central hypothesis that task-specific and well-designed prompt strategies are essential for effectively adapting LLMs to low-resource domains, with a particular focus on mental health applications.<br/><br/>This study systematically examines the role of prompt position across tasks and prompting methods, revealing that position significantly affects model performance and is often overlooked in prior work. A domain-specific prompting method is then explored to extract and summarise evidence of suicidal ideation with minimal supervision. Finally, a novel LLM-based framework is developed for multi-label life event annotation in the mental health domain, incorporating tailored prompting techniques and prompt position self-ordering to enable effective automatic annotation.
Degree
thesis:*- Name dc:type.qualificationname
- Ph.D.
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- University of Southampton
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Mao, Junyu
- Advisors dc:contributor.advisor
-
- Middleton, Stuart
- Niranjan, Mahesan
- ORCID person:identifier.orcid
Rights
- Language dc:language
- English, eng