Abstract
Real-world interactions involve a constant exchange of information between agents (be they humans or AIs) that are characterized by a limited availability of resources. These dialectical interactions, and the entailed properties, constitute the primary focus of this dissertation. They can be formalized by argumentative models of non-monotonic reasoning that provide real-world dialectical characterisations of arguments by resource-bounded agents. This thesis covers a wide range of implementations for such dialectical methods that span from proof theories to labelling algorithms, from argument schemes to dialogue protocols, and from explainable AI to decision support systems.The main contributions consist of: (1) the design of (sound and complete) dialectical argument game proof theories and (2) algorithmic procedures for the enumeration of dialectical labellings. (3) The formalisation of Explanation- Question-Response (EQR) protocols and specific EQR argument schemes yields dialogue specialized in conveying explanations. Upon these results, (4) the presentation of D-schemes, i.e., dialectical versions of EQR schemes, allows for EQR dialogue implementations capable of delivering explanations more suited to real-world resource-bounded agents. Finally, (5) a practical code-based implementation shows how a software chatbot can perform (a partial) EQR dialogue to assist users seeking information.
From this thesis stem different possible research directions, theoretical and practical. Focusing on the practical application, the herein automatization of the developed dialogue protocols allows the implementation of software capable of seamlessly enhancing clinical decision support systems by answering patients’ clarification needs. Furthermore, by implementing a fully-fledged EQR dialogue on the presented bot, it should be possible in the future to improve the software scope and enable a multi-purpose chatbot tailored to different explanation contexts and related functions.
| Date of Award | 1 Jun 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Peter McBurney (Supervisor) & Luca Viganò (Supervisor) |