Argument Mining into Active Learning Systematic Reviews: unlocking the synergy between MARGOT and ASReview

Active learning enhances the systematic review process by effectively screening a large amount of titles and abstracts, using machine learning in combination with human expertise. However, the intricacy of full-text (traditional) abstracts can lead to issues, such as token restrictions and longer processing time. In light of these challenges, this thesis harnesses the capabilities of argument mining to distill salient information from abstracts in order to refine the screening process. Therefore, I propose the integration between ASReview LAB, an active learning tool for systematic reviews, and MARGOT, an argumentation mining software. This suggested approach leverages the power of computational argumentation, illustrating its significant value in literature processing. On this basis, I conducted an experiment based on various benchmark data, employing machine learning techniques to extract features from both traditional and Argument Mined abstracts. These features informed subsequent classification models. Next, I test the consistency of the experiment and conduct a quantitative and qualitative analysis spotlighting the benefits of Argument Mined abstracts. Results indicate marginal differences between traditional and Argument Mined abstracts. Yet, in different scenarios, Argument Mined abstracts elevate the overall quality of the systematic reviews. Furthermore, the efficiency of machine learning models depends heavily on the intrinsic attributes of the data they process.