CLASP
The Centre for Linguistic Theory and Studies in Probability

Public defence of Amandine Decker's doctoral thesis

Thesis Title

A Monologue on Conversation: Characterising Dialogue across Interaction Types with a Focus on Topical Structure

Respondent:

Amandine Decker, Department of Philosophy, Linguistics and Theory of Science

Examining committee:

Docent Lucia Donatelli, Vrije Universiteit Amsterdam Professor David Traum, University of Southern California Docent Sharid Loáciga, University of Gothenburg

Opponent:

Professor Frederic Bechet, Aix Marsielle Université

Chair:

Researcher Bruno Guillaume, Université de Lorraine

Abstract

Dialogue constitutes a fundamental mode of human interaction and has become increasingly central in natural language processing with the growing use of conversational AI systems both in research and in practical applications. However, dialogue is a heterogeneous phenomenon encompassing a wide range of forms, arising from differences in participant configurations, interaction contexts, or communication media. This diversity creates challenges for the study of dialogue data, particularly in selecting suitable resources, defining the scope and limitations of our studies, and establishing meaningful evaluation methodologies. A principled characterisation of dialogue is therefore necessary to account for its variations and to support the design of robust computational approaches.

Among the phenomena that cut across different forms of dialogue, topical organisation plays a central role. The structure and evolution of topics shape conversational dynamics and are themselves influenced by the type of interaction being considered. Studying such phenomena requires moving beyond local conversational patterns and adopting a broader, macro-level perspective on dialogue, allowing different interacting processes and structures to be examined jointly.

This thesis investigates three main research questions: (1) how dialogue can be characterised in a way that accounts for its variability while informing the selection of suitable resources and evaluation methods; (2) whether topical structure can be automatically extracted from dialogue data, with a particular focus on open-ended multi-party conversations, that are more complex but also very common in real-world interactions; and (3) whether topical structure can be leveraged as a discriminative signal for differentiating dialogue family types. Through these questions, the thesis aims to contribute both analytical and methodological perspectives toward a broader understanding of dialogue structure and variation.

We answer these questions by proposing a principled characterisation of dialogue, CHIMERA — standing for CHaracterisation of dIalogue through Multidimensional Examination, Representation and Analysis — and by analysing the topical structure of dialogues across different interaction types, showing how CHIMERA can help determine the suitability of the resources employed and the scope of the analyses.

[Link to thesis]