|
Abstract:
|
Many people take part in online discussions, state their opinions on controversial topics, and bring forward their arguments. Sometimes, you ask yourself how similar your opinion is to the opinion of another participant of the discussion. For example, you might be reading the attitudes of political parties towards current political issues, and you ask yourself which party to vote for. But to undertake this kind of comparison, you need some means to calculate the similarities of attitudes in an argumentation. In this work, we delve into the issue of determining the similarities of individual views in an argumentation. We present a theoretical model to capture different opinions and arguments in argumentation contexts and develop a pseudometric for calculating the (dis)similarity between the views of two participants. Furthermore, we investigate how to make sure that such a distance measure yields intuitive results by looking at an empirical study where we collected human baseline results for argumentation similarity assessments. We propose different distance functions and study which best match human intuition and where the functions have limitations. Once we have the theoretical means to compare attitudes in argumentations, we examine two possible use cases. First, we explore how to achieve a clearer view in online discussion platforms with numerous arguments by pre-filtering arguments using neighborhood-based collaborative filtering. Our new argumentation platform deliberate includes such a filtering algorithm which uses our pseudometric for calculating the similarity of users based on their attitudes in the argumentation. We expound on results from an experiment with deliberate, where the influence of different filtering algorithms on the formation of opinion was studied. Moreover, we present our argumentation dataset for evaluating argument recommender systems which comprises several hundred user profiles. As a second use case, we introduce our argument-based Voting Advice Application (VAA) ArgVote, which computes the similarity of political views of parties and voters not only based on their opinion concerning central theses, but also considering their arguments. Although we could not demonstrate that our new matching algorithm based on our pseudometric was more accurate than the algorithm in classical VAAs, we were, nevertheless, able to show positive effects of our argument-based system on the understanding of political issues. The dataset containing the user profiles of our study participants is provided to improve the matching algorithms in future work. We subsequently present our idea for a VAA chat bot to address some issues with ArgVote which our experiment revealed. Our work lays the foundation for further exciting applications in the context of argumentations, for instance the clustering of voters, an automatic finding of compromises, or escaping filter bubbles. The impact of the developed methods, systems, and different user interfaces on opinion formation or political interest can be further researched in larger empirical studies.
|
