Argumentation Reasoning with Graph Isomorphism Networks for Reddit Conversation Analysis

Abstract

AbstractThe automated analysis of different trends in online debating forums is an interesting tool for sampling the agreement between citizens in different topics. In previous work, we have defined computational models to measure different values in these online debating forums. One component in these models has been the identification of the set of accepted posts by an argumentation problem that characterizes this accepted set through a particular argumentation acceptance semantics. A second component is the classification of posts into two groups: the ones that agree with the root post of the debate, and the ones that disagree with it. Once we compute the set of accepted posts, we compute the different measures we are interested to get from the debate, as functions defined over the bipartition of the posts and the set of accepted posts. In this work, we propose to explore the use of graph neural networks (GNNs), based on graph isomorphism networks, to solve the problem of computing these measures, using as input the debate tree, instead of using our previous argumentation reasoning system. We focus on the particular online debate forum Reddit, and on the computation of a measure of the polarization in the debate. We explore the use of two different approaches: one where a single GNN model computes directly the polarization of the debate, and another one where the polarization is computed using two different GNNs: the first one to compute the accepted posts of the debate, and the second one to compute the bipartition of the posts of the debate. Our results over a set of Reddit debates show that GNNs can be used to compute the polarization measure with an acceptable error, even if the number of layers of the network is bounded by a constant. We observed that the model based on a single GNN shows the lowest error, yet the one based on two GNNs has more flexibility to compute additional measures from the debates. We also compared the execution time of our GNN-based models with a previous approach based on a distributed algorithm for the computation of the accepted posts, and observed a better performance.