Advances in NLP have yet to lead to widespread automation in systematic reviews (SRs), which are complex and costly, ranging from USD 16-18 million annually. The review process involves searching for evidence, filtering relevant studies, appraising biases, extracting data, performing statistical analysis, and writing manuscripts.

Automation could reduce workloads and costs across the process. Although automatic study screening methods have been proposed, their real-world adoption is limited due to mismatches with existing workflows. Developing automatic information extraction methods that support study filtering and bias assessment remains a challenge, constrained by a lack of annotated datasets, static datasets, and varying data needs.

This thesis explores automation in citation screening, data extraction, and bias assessment. It evaluates active study screening methods for future scenarios, adapts weak supervision methods for diverse data types, and develops a resource to assess NLP techniques in bias assessment. These advances promise more efficient, transparent, and cost-effective SRs.