Multilingual Fake News Detection with Satire

Gaël Guibon (1) , Liana Ermakova , Hosni Seffih (2, 3) , Anton Firsov (4) , Guillaume Le Noé-Bienvenu (5)

1 LIS - Laboratoire d'Informatique et des Systèmes (LIS) (Marseille, Toulon)
2 UP8 - Université Paris 8 Vincennes-Saint-Denis
3 GEOLSemantics
4 Knoema Corporation Perm
5 UPN - Université Paris Nanterre

The information spread through the Web influences politics, stock markets, public health, people's reputation and brands. For these reasons, it is crucial to filter out false information. In this paper, we compare different automatic approaches for fake news detection based on statistical text analysis on the vaccination fake news dataset provided by the Storyzy company. Our CNN works better for discrimination of the larger classes (fake vs trusted) while the gradient boosting decision tree with feature stacking approach obtained better results for satire detection. We contribute by showing that efficient satire detection can be achieved using merged embeddings and a specific model, at the cost of larger classes. We also contribute by merging redundant information on purpose in order to better predict satire news from fake news and trusted news.

Mots clés: fake news; deception; artificial intelligence; machine learning; natural language processing; satire

Domaines: Méthodes et statistiques; Informatique et langage [cs.CL]; Machine Learning [stat.ML]

Gaël Guibon, Liana Ermakova, Hosni Seffih, Anton Firsov, Guillaume Le Noé-Bienvenu. Multilingual Fake News Detection with Satire. CICLing: International Conference on Computational Linguistics and Intelligent Text Processing, Apr 2019, La Rochelle, France. ⟨halshs-02391141⟩