OUTLINE

1.3

9

Outline

Figure 1.1. Overview of publications and how they relate to the chapters.

Figure 1.2. Visual Overview of the research questions and how they relate to the
chapters.

After the introductory Chapters 1 and 2, we continue with the publication-based
chapters that form the core of the thesis, which are structured in two parts.
Part I consists of a single chapter, Chapter 3, which presents a benchmarking
study of PUQ methods applied on real-world text classification datasets with
1-D convolutional neural networks and pretrained transformers. It motivates
a novel PUQ method, Deep Ensemble with Concrete Dropout, combining the
benefits of both methods, and showing promise for improving reliability and
robustness in NLP at a lower computational cost. The chapter concludes with
a discussion of the results, including targeted ablation studies, and provides
recommendations for future research.
Part II consists of three chapters, Chapters 4 to 6, which all focus on the more
applied research questions of realistic and efficient DU.