Executive Summary

Recent advancements in machine learning, particularly in natural language processing, have been marked by the emergence of large models pretrained on extensive datasets, contributing to breakthroughs in a variety of tasks. However, the increased complexity of these models has raised concerns about their opacity, limiting human understanding of decision-making processes. In response, this book focuses on methods for accountable and transparent machine learning, addressing concerns about adversarial manipulations, biases, privacy violations, and generalization issues. This chapter introduces accountability and transparency methods within the context of complex reasoning tasks over text, encompassing domains such as fact checking, question answering, and natural language inference. The included papers, detailed in Sect. 1.2, collectively contribute to the understanding and improvement of machine learning models in these domains. Notably, the research explores vulnerabilities, develops challenge datasets, and proposes novel approaches for generating natural language explanations, thereby enhancing the accountability and transparency of machine learning models.