Open Access

Mostafa Mahmoud Amin

• Foundation models, particularly LLM like ChatGPT and its underlying GPT models, have transformed the landscape of NLP and offer new avenues for affective computing. This dissertation explores the integration of the GPT foundation models with affective computing tasks to enhance the bridging between human emotions and advanced AI technologies. This thesis contributes to the field by introducing the "prompting paradigm", a novel computational framework that leverages foundation models to solve affective computing problems; this paradigm provides the skeleton for the main studies underlying this work and the basis for future foundation models approaches to affective computing. Key findings of this work indicate that LLM exhibit emergent affective computing capabilities, offering nuanced emotional insights and improved task performance. A follow-up study is conducted to explore a wide range of affective computing tasks, grouped into three main categories: sentiment-based problems,Foundation models, particularly LLM like ChatGPT and its underlying GPT models, have transformed the landscape of NLP and offer new avenues for affective computing. This dissertation explores the integration of the GPT foundation models with affective computing tasks to enhance the bridging between human emotions and advanced AI technologies. This thesis contributes to the field by introducing the "prompting paradigm", a novel computational framework that leverages foundation models to solve affective computing problems; this paradigm provides the skeleton for the main studies underlying this work and the basis for future foundation models approaches to affective computing. Key findings of this work indicate that LLM exhibit emergent affective computing capabilities, offering nuanced emotional insights and improved task performance. A follow-up study is conducted to explore a wide range of affective computing tasks, grouped into three main categories: sentiment-based problems, psychology-based problems, and problems with implicit signals. GPT models, including GPT-3.5 and GPT-4, show the most superior performance on psychology-based problems, compared to specialised models. Furthermore, GPT models have shown strong performance in sentiment-based problems; however, specialised NLP methods still exhibit the most superior performance in many cases. Finally, GPT models perform poorly on problems with implicit signals. The GPT models are compared against NLP baselines, featuring different generations of NLP methods, namely using Bag-of-Words, Word2Vec, and RoBERTa features. The sentiment-based problems are sentiment analysis, opinion extraction, aspect extraction, aspect target polarity classification, subjectivity detection, sentiment intensity ranking, and emotions intensity ranking. The psychology-based problems are toxicity detection, suicide-tendency detection, and well-being and stress detection. The implicit signals problems are sarcasm detection, personality assessment, and engagement measurement. The nuances of the paradigm are also explored in other studies. One of the studies conducts a comprehensive analysis of prompt sensitivity and sampling strategies, highlighting the importance of precise prompt engineering to optimise model outputs. Key prompting insights are using conservative sampling, and instructing the model to behave as an expert or to conduct step-by-step problem solving before giving a final answer, known as CoT. Furthermore, novel fusion strategies that combine outputs from LLM with traditional NLP methods demonstrate performance improvements. A key idea to employ this is using verbose responses from LLM, then further analysing them with NLP methods instead of simple text parsing. This assists in acquiring scores that can be fused easily with traditional NLP methods. Furthermore, ethical considerations and compliance with regulatory frameworks are addressed to ensure the responsible deployment of AI technologies in affective computing applications, in alignment with the European Union AI Act. By exploring the intersection of foundation models and affective computing, this research lays the groundwork for future studies, focusing on refining methodologies for the use of foundation models within affective computing. This assists in maximising the impact of affective computing on real-world applications while considering societal impacts.…