Explanation-Based Scientific Natural Language Inference

Abstract

Building systems that can explain and understand the world is a long-standing goal for Artificial Intelligence (AI). The ability to explain, in fact, constitutes an archetypal feature of human rationality, underpinning communication, learning, and generalisation, as well as one of the mediums enabling scientific discovery through the formulation of explanatory theories. As part of this long-term goal for AI, a large body of research in Natural Language Processing (NLP) focuses on the development and evaluation of explanation-based inference models, capable of reasoning through the interpretation and generation of natural language explanations. However, research in Explanation-based Natural Language Inference (NLI) presents several fundamental challenges. Firstly, the applied methodologies are still poorly informed by theories and accounts of explanations. Current work, in fact, rarely recur to formal characterisations of the nature and function of explanations, and are limited to generic explanatory properties. This gap between theory and practice poses the risk of slowing down progress in the field, missing the opportunity to formulate clearer hypotheses on inferential properties of explanations and well-defined evaluation methodologies. Secondly, Explanation-based NLI models still lack robustness and scalability for real-world applications. In particular, existing approaches suffer from several limitations when it comes to composing explanatory inference chains from large facts banks and performing abstraction for NLI in complex domains. This thesis focuses on scientific explanation as a rich theoretical and experimental framework for advancing research in Explanation-based NLI. In particular, the goal of the thesis is to investigate some of the fundamental challenges in the field from both a theoretical and an empirical perspective, attempting to derive a grounded epistemological-linguistic characterisation to inform the construction of more accurate and scalable Explanation-based NLI models in the scientific domain. Overall, the research described in the thesis can be summarised in the following scientific contributions: 1. An extensive study on the notion of a scientific explanation from both a categorical and a corpus-based perspective aimed at deriving a grounded characterisation for explanation-based NLI. The study reveals that explanations cannot be entirely defined in terms of inductive or deductive arguments as their main function is to perform unification, fitting the event to be explained into a broader underlying regularity. Moreover, the study suggests that unification is an intrinsic property of existing corpora, emerging as explicit and recurring explanatory patterns in natural language. 2. A novel computational model based on the notion of explanatory power as defined in the unificationist account of scientific explanation. Specifically, the model can be adopted to capture explicit explanatory patterns emerging in corpora of natural language explanations and flexibly integrated into explanation-based NLI architectures for downstream inference tasks. 3. An empirical study on the impact of the explanatory power model on explanation-based NLI in the scientific domain, integrating it within sparse, dense and hybrid architectures, and performing a comprehensive evaluation in terms of inferential properties, accuracy and scalability. 4. A systematic evaluation methodology to inspect and verify the logical properties of explanation-supporting corpora and benchmarks. The study, aimed at providing a critical quality assessment of gold standards for NLI, reveals that a majority of human-annotated explanations represent invalid arguments, ranging from being incomplete to containing identifiable logical errors.

Date of Award	31 Dec 2022
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Andre Freitas (Supervisor) & Tingting Mu (Supervisor)