Modelling Explanatory Reasoning: A Study on Explanatory Patterns and Graph-Based Inference

Abstract

The ability to generate meaningful explanations is a crucial component in advancing artificial intelligence (AI), especially in systems that aim to aid decision-making in complex domains. While deep learning models have achieved remarkable success in many tasks, they often suffer from the limitation of learning only surface-level patterns, such as annotation artefacts and biases. These models lack the deeper cognitive ability to reason abstractly, draw analogies, and unify information across different layers of abstraction. Recent research emphasizes that overcoming this limitation requires a shift towards reasoning frameworks grounded in unification, analogy, and abstraction. Constructing explanations, particularly in AI, involves retrieving facts and connecting them across different levels of abstraction, from word-level to sentence-level to conceptual-level. This hierarchical structure demands models that can navigate and unify information within a framework of relationships, causal patterns, and abstract reasoning. Graph-based methods present a promising solution for encoding these explanatory patterns, as they naturally model complex relational structures and allow for both fine-grained and high-level abstraction in explanation generation. This paper addresses several critical research questions in the context of explanation generation through the lens of epistemological and linguistic perspectives. It first explores the explanation patterns identified in the literature, particularly focusing on unification, causality, and the hierarchical structure of explanations that facilitate abstraction. The core investigation then centres on whether graph-based deep learning methods can effectively encode these explanatory patterns and how variations in the graph structure and neural network design impact their performance. Building upon this foundation, we propose strategies for constructing improved graph structures and graph convolution networks, combining insights from both theoretical and empirical investigations. In particular, we introduce the Hierarchical Hypergraph Attention Network (HHGAT) and conduct experiments to assess its potential to enhance the encoding and retrieval of explanatory patterns, aiming to advance the capabilities of AI systems in providing interpretable and coherent explanations.

Date of Award	19 Mar 2025
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Andre Freitas (Main Supervisor) & Lucas Cordeiro (Co Supervisor)