Argument Browser - A Tool for Pluralistic Discourse

Abstract

Public discourse in digital spaces, and in particular in the comment section of news articles, presents clear shortcomings. The quantity of information is overwhelming, the most heard voices tend to be highly polarised and different opinions incline to cluster themselves, resulting in echo-chambers and one-sided discussions. To tackle these challenges, we propose a tool to foster constructive, diversified and inclusive arguments. The tool classifies these comments and allows the user to filter and visualise them along different categories, which include stance, political view, topic, etc.

Our pipeline uses state of the art language models to automatically label each comment in the desired categories. Because there are no available datasets containing labelled comments and the manual labelling process is complex, this work focuses also on lowering the number of needed labelled training entries. Semi-supervised learning algorithms are employed and improved thanks to a novel text data augmentation technique, which, with comparable classifying performance, is 25 times faster than the current standard. Moreover, with the same objective of lowering the number of labelled entries, an inductive transfer learning approach is designed and added to the pipeline, in order to transfer knowledge from datasets similar to the target one. The combination of the proposed techniques can lead up to 38% performance improvement in case of low number of labelled data points when compared to employing only supervised learning classification techniques. As last step, the proposed approach is shown to work on a dataset extracted from the comment section of an English newspaper and manually labelled by us.