Abstract
In recent years, the massive use of social networks and the large number of usersinvolved has been an essential source for studies and analyses, on a variety of tasks.Controversy on social media is one such task studied in the scientific literature. Contro-versial topics are defined as subjects where there is widespread disagreement betweenusers on a specific event, showing a wide divergence of opinions. They are often repre-sented by sometimes highly polemical debates or contestations around sensitive topicsor statements that do not meet with unanimous approval. The aim of this project is tostudy controversy as a whole on social networks, by investigating both the textual andstructural properties of these topics, based on recent deep learning methods.The advent of machine learning, and more specifically deep learning methods,has led to improved representation of data of all types, using more complex learningtechniques. In this thesis, we focus on 2 types of domain in particular :• Natural language processing (NLP), in order to learn complex textual represen-tation using deep models based on attention mechanisms, such as BERT.• Graph processing, in order to learn how to represent unstructured data such asgraphs. To this end, we are studying the progress made in graph neural networks(GNNs), and the methods derived from them.In this thesis, we will present through four parts : (1) the state of the art in contro-versy analysis methods and graph neural networks, (2) a method to explain controver-sial topics by analyzing content of Twitter communities, (3) a graph and text basedmethod to quantify controversial topics on Twitter, based on user polarization aroundcommunities. (4) a method for early prediction of controversial posts from Redditdiscussion between users.Keywords – Machine learning, Graph neural networks, Natural language processing, auto-matic detection, Controversy, Explainability, Social media