• PhD Student:
  • Caren Al Anaissy
  • Funding : ANR, Région HdF
  • PhD defended on :
  • Jul 4, 2024

Computational argumentation allows for evaluating and assessing the acceptability of arguments in order to infer conclusions out of the reasoning process or to resolve conflicts of opinions in dialogues. This thesis introduces several new developments in computational argumentation, building on the foundations of Dung’s abstract argumentation framework.

First, the thesis introduces an analysis of seven types of semantics for bipolar argumentation frameworks, extending Dung’s abstract argumentation framework by incorporating diverse interpretations of support alongside attack. We study three groups of semantics: defense-based, selection-based, and reduction-based with respect to ten principles. This principle-based analysis results in an extensive evaluation of bipolar argumentation semantics, reflecting the depth and breadth of the study.

Second, the thesis refines an existing impact measure and defines a novel one for gradual argumentation semantics to quantify the influence of argument sets on the acceptability of other arguments. Several principles are defined to evaluate these impact measures, which are essential for demonstrating the effect of specific arguments in strengthening or weakening an argument’s acceptability.

Third, neural networks, including advanced language models, have achieved significant advancements yet continue to face challenges in reasoning. Considering that argumentation is a fundamental aspect of human thinking and decision-making, this thesis investigates whether neural networks can learn argumentation semantics, serving as an important step toward developing more realistic reasoning systems. This includes a practical application to real-world datasets—Twelve Angry Men and Debatepedia—utilizing three neural network architectures: Multilayer Perceptron, Graph Convolution Network, and Graph Attention Network. The results demonstrate that these architectures, particularly the Graph Convolution Network, can effectively capture bipolar gradual argumentation semantics, highlighting the importance of using graph-based neural architectures that model argumentation graphs explicitly.

Finally, the thesis shows how to use argumentation to enhance the explainability and understandability of the persuasion mechanisms and techniques conducted by persuasive chatbots. It surveys the existing legal framework and ethical principles, and it proposes a graphical representation of argumentative interactions to improve system transparency and user trust. This approach not only helps in making persuasive chatbots more understandable but also allows for user feedback on the ethical dimensions of arguments.

Composition of the committee:

Supervisors:

  • Srdjan Vesic, CRIL, CNRS - Artois University

  • Nathalie Nevejans, CDEP - Artois University

Reviewers:

  • Leila Amgoud, IRIT, CNRS - Paul Sabatier University

  • Anthony Hunter, University College London

Examiners:

  • Sébastien Konieczny, CRIL, CNRS - Artois University

  • Nico Potyka, Cardiff University