Understanding and Enhancing the Explainability of Neural Natural Language Inference Models

Abstract

The rapid advancements in deep learning have led to state-of-the-art performance in numerous natural language processing (NLP) tasks. However, the complexity and opacity of these neural models pose significant challenges for Explainable Artificial Intelligence (XAI). Research has demonstrated that biases and artifacts can spuriously inflate model performance but undermine their reliability and applicability. Natural language inference (NLI) is a central problem in natural language understanding and artificial intelligence, aiming to model logical relationships between two sentences. This dissertation is centred around improving the explainability of neural NLI models via different methods. Specifically, we explore solutions for four key problems. The first is designing inherently explainable frameworks whose underlying mechanism naturally provides explainability. For this purpose, we propose to integrate natural logic, a classic logic framework for NLP, with neural networks. Such a neural-symbolic framework combines the representational power of neural networks with the explainability of formal methods. The second problem is designing NLI models that can handle complex reasoning that involves diverse reasoning aspects. To address this challenge, we propose a Mixture of Prompt Experts framework, where each expert is specialized to model a particular reasoning perspective. We hope to enhance reasoning performance and trace how individual modules contribute to the final decision, thereby enhancing explainability. In addition to the above approaches that focus on building white-box models, the third challenge is identifying black-box NLI models' behaviour and weaknesses where the model architectures are inaccessible. We focus on adversarial attack techniques to uncover the models' vulnerabilities and observe their behaviours. By analyzing how a black-box model reacts to carefully crafted adversarial examples, we can infer patterns in its decision-making behaviour, shedding light on its explainability. While the above studies have been conducted in the classic setup of determining logic relationships between one premise and one hypothesis, the last key problem arises from the complex NLI task where multiple premises are used to prove a hypothesis. To tackle this issue, we propose a framework incorporating structure-aware modules to explicitly model reasoning steps. By revealing each reasoning step and the corresponding search process, the framework naturally provides reasoning explainability.