Название: Accountable and Explainable Methods for Complex Reasoning over Text
Автор: Pepa Atanasova
Издательство: Springer
Год: 2024
Страниц: 208
Язык: английский
Формат: pdf (true)
Размер: 26.7 MB

This thesis presents research that expands the collective knowledge in the areas of accountability and transparency of Machine Learning (ML) models developed for complex reasoning tasks over text. In particular, the presented results facilitate the analysis of the reasons behind the outputs of ML models and assist in detecting and correcting for potential harms. It presents two new methods for accountable ML models; advances the state of the art with methods generating textual explanations that are further improved to be fluent, easy to read, and to contain logically connected multi-chain arguments; and makes substantial contributions in the area of diagnostics for explainability approaches. All results are empirically tested on complex reasoning tasks over text, including fact checking, question answering, and natural language inference.

A major concern with Machine Learning (ML) models is their opacity. They are deployed in an increasing number of applications where they often operate as black boxes that do not provide explanations for their predictions. Among others, the potential harms associated with a lack of understanding of the models’ rationales include privacy violations, adversarial manipulations, and unfair discrimination. As a result, the accountability and transparency of ML models have been posed as critical desiderata by works in policy and law, philosophy, and Computer Science.

In Computer Science, the decision-making process of ML models has been studied by developing accountability and transparency methods. Accountability methods, such as adversarial attacks and diagnostic datasets, expose vulnerabilities in ML models that could lead to malicious manipulations or systematic faults in their predictions. Transparency methods explain the rationales behind models’ predictions, gaining the trust of relevant stakeholders and potentially uncovering mistakes and unfairness in models’ decisions. To this end, transparency methods have to meet accountability requirements as well, e.g., being robust and faithful to the underlying rationales of a model.

This book presents my research that expands our collective knowledge in the areas of accountability and transparency of ML models developed for complex reasoning tasks over text. First, this book contributes two methods for accountable ML models. They generate adversarial inputs and a diagnostic dataset demonstrating significant model vulnerabilities and suggesting ways to correct those. In the area of transparency of ML models, this book advances state-of-the-art methods for generating textual explanations that are further improved to be fluent and easy to read as well as to contain logically connected multi-chain arguments. Finally, this book makes contributions in the area of diagnostics for explainability approaches with a set of properties for evaluating existing explainability techniques and methods for enhancing those further in produced explanations. All of the contributions are empirically tested on complex reasoning tasks over text, including fact checking, question answering, and Natural Language Inference (NLP).