How to Read Less: On the Benefit of Human-in-the-loop Incremental Learning for Systematic Literature Review

Systematic literature reviews (SLRs) are the primary method for aggregating and synthesizing evidence in evidence-based software engineering (SE). Primary study selection is a critical and time-consuming SLR step in which reviewers use titles, abstracts, or even full texts to evaluate thousands of studies to find the dozens of them that are relevant to the research questions. We seek to reduce the effort of primary study selection in SE SLRs by exploring and refactoring the state-of-the-art human-in-the-loop incremental learning techniques from evidence-based medicine and legal electronic discovery. By refactoring those methods, we discovered FASTREAD, which is a new state-of-the-art in human-in-the-loop incremental learning for SE SLRs. Tested on two data sets generated from existing SE SLRs of Hall, Wahono, et al., FASTREAD outperforms the current state-of-the-art methods. Our results suggest that FASTREAD is able to find $90\%$ of the studies found by standard manual methods, by only reading less than $10\%$ of the candidate studies.