Neural argumentation networks (NANs) are artificial neural networks that learn according to abstract argumentation semantics. When pro- vided with an input data set of argument acceptability data, NANs learn a solution defeat relation that is consistent with the input. The objective of this dissertation is to describe and evaluate the theory and application of translating complete argumentation semantics to NAN architectures for effective defeat relation learning. The performance of algorithms designed to address this objective are evaluated against two distinct forms of data input and two distinct solution conditions. Two distinct data categories are required to address the alternative approaches for extension-based (2-valued) and labelling-based (3-valued) acceptability data. A further distinction is made between evaluating effectiveness at finding a solution defeat relation that can perfectly express a given input of argument acceptability data, defined as the σ-consistency problem, and the more general noisy case where perfect expression is not expected. It is within the scope of these distinctions that the NANs are crafted, with accompanying learning algorithms that are evaluated theoretically based on time efficiency and empirically based on the fidelity of the output defeat relation to the ground truth defeat relation that exists between a given set of arguments. The analysis shows that algorithms designed for 3-valued data were significantly more effective, in terms of defeat relation fidelity, than algorithms designed for 2-valued data. The 3-valued data algorithms were able to attain exceptionally high defeat relation fidelity even when provided relatively small sets of data compared with contemporary machine learning requirements of ‘Big Data’.
Learning How to Defeat Arguments Using Data-Driven Neural Argumentation Networks
Mumford, J. (Author). 1 Jul 2022
Student thesis: Doctoral Thesis › Doctor of Philosophy