A Neural Approach for Incomplete Argumentation Frameworks

Abstract

Incomplete Argumentation Frameworks (IAFs) provide an intuitive modelling of argumentative scenarios where uncertainty exists regarding which arguments and attacks will be used in a debate. A significant limitation of expressive frameworks like IAFs is their high theoretical complexity, which renders exact reasoning algorithms impractical in applications under time constraints or in interactive settings involving human users. For this reason, we want to study the efficacy of approximate algorithms that address the same reasoning problems but with significantly reduced execution times. To achieve this, we have defined a Neural Network, using the Edge-Featured Graph Attention Network (EGAT) architecture, which is trained to solve classical acceptability problems of IAFs. We have conducted an empirical evaluation of our novel approach which shows that our EGAT-based approach provides very satisfying results w.r.t. standard metrics like accuracy, precision, recall and the F1-score. This research paves the way for the application of Neural Networks for efficiently solving reasoning tasks in expressive generalizations of Dung’s abstract argumentation.