The full text of this thesis is available from the Newcastle University Library website: https://theses.ncl.ac.uk/dspace/handle/10443/791
Shearer, J., School of Computing Science, University of Newcastle upon Tyne
Realism for embodied conversational agents (ECAs) requires both visual and behavioural fidelity. One significant area of ECA behaviour, that has to date received little attention, is non-verbal behaviour. Non-verbal behaviour occurs continually in all human-human interactions, and has been shown to be highly important in those interactions. Previous research has demonstrated that people treat media (and therefore ECAs) as real people, and so non-verbal behaviour is also important in the development of ECAs. ECAs that use non-verbal behaviour when interacting with humans or other ECAs will be more realistic, more engaging, and have higher social influence. This thesis gives an in-depth view of non-verbal behaviour in humans followed by an exploration of the potential social influence of ECAs using a novel Wizard of Oz style approach of synthetic ECAs. It is shown that ECAs have the potential to have no less social influence (as measured using a direct measure of behaviour change) than real people and also that it is important that ECAs have visual feedback on their interactants for this social influence to maximised. Throughout this thesis there is a focus on empirical evaluation of ECAs, both as a validation tool and also to provide directions for future research and development. Present ECAs frequently incorporate some form of non-verbal behaviour, but this is quite limited and more importantly not connected strongly to the behaviour of a human interactant. This interactional aspect of non-verbal behaviour is important in human-human interactions and results from the study of the persuasive potential of ECAs support this fact mapping onto human-ECA interactions. The challenges in creating non-verbally interactive ECAs are introduced and by drawing corollaries with robotics control systems development behaviour-based architectures are presented as a solution towards these challenges, and implemented in a prototypical ECA. Evaluation of this ECA using the methodology used previously in this thesis demonstrates that an ECA with non-verbal behaviour that responds to its interactant is rated more positively than an ECA that does not, indicating that directly measurable social influences will be possible with further development.