Posterior Probabilities, Active Learning, and Transfer Learning in Technology-Assisted Review

Technology-Assisted Review (TAR) refers to the human-in-the-loop machine learning process whose goal is that of maximizing the cost-effectiveness of a review (i.e., the task of labeling items to satisfy an information need). This thesis explores and thoroughly analyzes: the applicability of the SLD algorithm to TAR scenarios; the usage of active learning combined with the MINECORE framework, effectively improving the framework performance; the portability of machine/deep learning models for the production of systematic reviews in empirical medicine. Finally, the thesis proposes a new algorithm, based on SLD, called SALt, which improves the class prevalence estimates on active learning scenarios, with respect to the current state-of-the-art.