Personalisation in Dialogue Systems: From Accessible Interaction to Pragmatic Awareness

The thesis focuses on the personalisation of natural language technologies, particularly in dialogue systems, with the goal of addressing user-specific needs. The work aims to explore how personalised systems can better serve individual users by improving both the clarity of communication and accessibility, with a specific emphasis on users with disabilities. This doctoral research investigates the gap between current state-of-the-art language technologies and the real-world needs of users, asking how dialogue systems can be adapted to respond more effectively to diverse user characteristics, such as sociodemographic factors. The broader objective is to determine whether personalising dialogue interactions based on these factors can lead to more natural, effective, and satisfying communication experiences. The work develops in two distinct but related directions. The first focuses on how dialogue systems can adapt to the needs of users with visual disabilities, particularly through accessible dialogic interaction. A key goal is to understand how such systems can be designed to make complex, visually oriented graphs, especially in fields like computer science, mathematics, and physics, more accessible. The study explores how these systems can not only improve communication, but also act as tools to overcome the barriers that visually impaired people face when accessing digital content. The results show that visually impaired people preferred dialogic interaction for graph exploration and that dialogic access may support efficient information retrieval and understanding. Moreover, the thesis explores how this interaction can be made more controlled yet dynamic by leveraging large language models, enabling more adaptive and personalised exchanges. The second direction concerns the study of irony as a key phenomenon for pragmatic personalisation in dialogue. Irony poses a major challenge for natural language systems, as it depends on subtle contextual cues, shared knowledge, and user-specific interpretation. This doctoral research investigates how large language models can be adapted to recognise and generate irony appropriately, depending on user characteristics and conversational context. The results indicate that personalised irony detection and generation models outperform general-purpose ones, particularly in recognising user-dependent pragmatic cues. In conclusion, this doctoral thesis aims to advance the understanding of how personalisation in dialogue systems can be harnessed to meet diverse user needs, with a particular emphasis on accessibility and inclusivity. By combining research on adaptive interaction and pragmatic modelling, it contributes to the development of dialogue systems that are not only more intelligent, but also more human-centred