Responsiveness: A Feature for Predicting the Productivity of Non-Convergent Discussions

This work focuses on estimating the impact of produced arguments on the discussion's upcoming development. We sought to define a feature upon which the productivity of non-convergent discussions can be predicted. We rely on the Bakhtinian notion of responsiveness, the degree to which a speaker embeds or builds on the arguments of an interlocutor. We demonstrate the potential of this feature, using a corpus of 10,000 threads extracted from Reddit's 'Change My View' forum. Utilizing Synthetic Minority Oversampling Technique, we experimented with several supervised machine learning algorithms, each of which drew on a different set of selected features. The performance of the obtained models was evaluated through repeated stratified 10-fold cross-validation. Our preliminary results are encouraging: responsiveness contributes to accurate prediction of discussion productivity.