Personalised dialogue generation involves producing responses following a conversational context, specifically reflecting an agent's predefined persona characteristics when necessary. However, this task faces significant challenges due to limited persona-specific data and the complexity of adapting large pre-trained models effectively and efficiently to diverse behavioural styles.

This thesis proposes several novel adaptation methods, including both architectural modifications and parameter-efficient techniques, to overcome these challenges. First, the Persona-Adaptive Attention (PAA) mechanism dynamically balances persona and dialogue context information, improving consistency in generated responses. Second, the LAPDOG framework introduces retrieval-augmented generation, focusing on learning how to optimise retrieval augmentation through joint training of the retriever and generator, using external stories to enrich persona data and improve response quality. Third, Selective Prompt Tuning (SPT) employs a trainable dense retriever to dynamically select contextually appropriate soft prompts for large language models, thereby enhancing the diversity and relevance of personalised responses. Finally, Hadamard High-Rank Adaptation (HiRA) addresses expressiveness limitations in traditional low-rank adaptation methods by employing Hadamard product decomposition, maintaining computational efficiency and yielding performance gains on the personalised dialogue generation task.

Collectively, these approaches significantly advance personalised dialogue generation by addressing core challenges such as limited persona data, context-persona balancing, prompt selection, and expressiveness of adaptation. They enhance response coherence, adaptability, and quality while maintaining computational efficiency. These contributions underscore the potential of combining architectural and parameter-efficient methods to build scalable and effective personalised conversational systems.