Machine Learning and Open Science: On Risks and Challenges

Recent years have witnessed substantial growth in Machine Learning (ML) and Natural Language Processing (NLP), largely fueled by the accessibility and openness of data and models—a cornerstone of Open Science. This dissertation builds on this foundation by integrating additional principles of Open Science—transparency, scrutiny, critique, and reproducibility—into the study of these fields. The dissertation extensively explores and tackles the challenges in reproducibility across both automatic and human evaluations in ML. It begins by unraveling the hidden complexities in evaluating uncertainty, emphasizing the necessity of rigorous statistical analysis, which include effect size and power analysis, and acknowledges the persistent risks of false discoveries despite careful considerations. This is complemented by a comprehensive guide to conducting and reporting uncertainties in evaluations, presenting a crucial resource for researchers to enhance the reliability of their findings. Further dissecting reproducibility challenges, we investigate the trends in availability of research artifacts and examines the impact of community-driven initiatives aimed at improving reporting practices. Furthermore, we present reproducibility assessment of eight scientific papers. Despite certain improvements spurred by community-driven initiatives for better reporting practices, there remain major issues that hinder reproducibility. An in-depth case study on the reproducibility of a text simplification pipeline reveals several overlooked reproducibility challenges such as bugs and dependency issues. Reproducibility of human evaluations is also scrutinized through two case studies. After observing mixed results, we identify several factors that contribute to inconsistencies in human evaluations, including small sample sizes and dynamic conditions. Through these analyses, the dissertation underscores the ongoing challenges in achieving reproducibility in ML and NLP, offering insights to bolster the reliability of future research within these dynamic fields.