Finding the needle in the haystack with Active Learning – Gerbrich Ferdinands
Scholars are confronted with ever-larger amounts of textual data. All this data present new and unique opportunities to scholars, while simultaneously confronting them with unprecedented challenges. How to select relevant text effectively and efficiently from an almost unlimited amount of data? Conducting a systematic review on this data is often a very time consuming and tedious task. Reviewers have to manually scan thousands of abstracts of scientific articles and assess their relevance to the research question at hand. For an experienced reviewer, it takes between 30 seconds to a couple of minutes to classify a single abstract, which easily results in hundreds of hours spent merely on abstract screening. In this day and age when the field of artificial intelligence is thriving at an unprecedented pace, one would imagine that this large amount of manual work can be reduced or even completely replaced by some smart machine learning software.
The AI-project we are working on, titled ASReview, includes the newest insights from active learning. Our software saves researchers time by decreasing the number of papers that need to be screened by the researcher. Machine learning algorithms ensure that (almost) all relevant papers are shown to the researcher nonetheless. Replications of existing systematic reviews have been performed by ASReview, with promising first results.