A text and GNN based controversy detection method on social media

Expressed opinions on social media frequently cause a controversy. Controversial content refers to content that attracts different opinions and interrogations, implying interaction between communities. Its automatic identification remains a challenging task. Most of the existing approaches rely on the graph structure of discussion and/or the content of messages but did not deeply explore the recent advances on Graph Neural Network ( gnn ) to predict if a discussion is controversial or not. This paper aims to combine both user interactions present in the graph structure of a discussion and the discussion text features to detect controversy. We rely on sampling techniques to reduce the size of large graphs and augment the graph training set if needed. Our proposed approach relies then on gnn techniques to encode the initial (or sampled) graph in an embedding vector before performing a graph classification task. We propose two controversy detection strategies. The first one is based on a hierarchical graph representation learning to take advantage of hierarchical relationships that could exist between users. The second one is based on the attention mechanism, which allows each user node to give more or less importance to its neighbors when computing node embeddings. We present different experiments conducted with data sources collected from both Reddit and Twitter to show the applicability of our approach to different social networks. Conducted experiments show the positive impact of combining textual features and structural information in terms of performance and accuracy.