Faking news recommendation : an exploration of content representation

Algorithms are making their way into many disciplines, this paper studies their use in journalism. We focus on the representation of text content for recommendation. We introduce a new text clustering technique based on graph modularity and inter-article similarities that outperforms classical clustering methods. We then explore the case of adversarial examples on text classifiers, writing manually articles to be incorrectly classified by the best performing algorithms on our dataset. We observe differences in difficulty according to the type of classifier.