Argumentation Frameworks in the Wild : An Empirical Study of Argumentation Graph Properties

Sammanfattning

This thesis does an empirical analysis of explicit generated argumentation graphs from a real-world dataset. Thereby, the thesis expands existing research, developing a procedure to extract and analyze the underlying structure of generatedargumentation graphs and applying this procedure to a real-world dataset based on movie ratings. Studying the structure of argumentation graphs generated from real data allows for the validation and extension of theory-oriented work on computational argumentation. The graphs studied in the thesis are based on existing work in which computational argumentation is applied to a recommendation system problem. This thesis looks into the underlying structure of 23,210 generated graphs and investigates assumptions made in theory. The thesis finds that the dataset used to generate the graphs generates un-balanced graphs where some parts of the graphs can be empty. It accordingly finds that the conceptual design of the argumentation-based system should be expanded to take this into account when weighting influence factors during the generating of argumentation graphs. The empty parts of the graphs occur because the underlying structure is relying on the distribution of common elements between movies that are by themselves unevenly distributed. These structuresare genre, actors, and directors. Many movies have genres in common while some have common actors but a significant number of movies lack common directors leading to empty sub-graphs.