Multimodal approach by embedding text and graphs for the detection of abusive messages

Abusive behaviors are common on online social networks, forcing hosts of such platforms to find new moderation tools. Various methods based on the textual content or the structure of the conversation have thus emerged. Furthermore, new generic embedding methods have been proposed to represent text and graphs. Those approaches benefit the current exponential growth in available data and computing power. In this work, we evaluate five text embedding methods and four graph embedding methods on an abusive message detection task. These two types of embedding are not based on the same information, therefore we also study various combinations of these methods. Our results are comparable, and even better for text, to standard approaches based on feature engineering. The combination of text and graph embeddings finally brings a clear improvement in performance.