Crafting neural argumentation networks - learning argument defeat relations from acceptability data

2^nd June 2021, 11:00
Jack Mumford

Abstract

In this presentation I will present an overview of my PhD research in which connectionist architectures were developed, Neural Argumentation Networks (NANs), that learn according to abstract argumentation semantics. The objective is to address the inverse argumentation problem in which a solution defeat relation is required, given inputs of argument acceptability labellings.



But why would anyone care about this? There is ample debate about the relative merits of symbolic vs connectionist methodologies. Briefly put, my research was motivated by the intuition that a connectionist architecture employing argumentation semantics may offer a route to XAI-friendly decisions via argumentation whilst benefitting from the data-driven connectionist approach that avoids the computational bottleneck of relying on manual construction.



How does this work? In traditional abstract argumentation, one has three ingredients: arguments, an argument relation, and acceptability labellings. In order to derive the acceptability labellings, one must already have the argument relation. This has historically been manually derived through expert (or perhaps not so expert) judgement. An alternative approach is to deploy ML techniques to learn the argument relation from data. I will present one approach to achieving this objective, and assess the complexity and performance of several distinct learning algorithms designed for the NAN architecture, offering a comparison to a benchmark taken from an argumentation synthesis implementation.

Biography

Jack is unsure whether he is supposed to boast about himself here. He has decided to play it safe and simply work backwards in time. Before joining Professor Katie Atkinson as a Research Associate, he pursued his PhD at King's College London (KCL) with Professor Simon Parsons, Dr Elizabeth Black and Dr Isabel Sassoon. He also studied at KCL for an MSc in Intelligent Systems (which has since been, perhaps inevitably, rebranded to the MSc in Artificial Intelligence), and has a BSc in Mathematics from the Open University. He also has a BA in English Literature from Aberystwyth University that he tends to neglect to mention.