Detecting argumentative discourse in online chat experiments

This paper applies argument mining techniques to chat data obtained from an online survey experiment with deliberative content. We investigate the task of automatically detecting chat messages that give justification for an underlying claim. We use bag-of-words features as well as state of the art wordand sentence-embedding models to train different classifiers on the given task. In contrast to previous studies, our results indicate that structural features are less important to predict argumentative reasoning in the chat. Moreover, the random forest classifier with features extracted from BERT has the best overall performance in the classification task. Our results show that argument mining techniques can be successfully applied to chat data obtained from economic experiments. It offers the chance to answer empirical research questions such as the effect of deliberation on economic, social, or political behaviour.