IRIT at e-Risk 2018

Faneva Ramiandrisoa (1) , Josiane Mothe (1, 2) , Farah Benamara (3, 4) , Véronique Moriceau (3, 5)

1 IRIT-SIG - Systèmes d’Informations Généralisées
2 UT2J - Université Toulouse - Jean Jaurès
3 IRIT-MELODI - MEthodes et ingénierie des Langues, des Ontologies et du DIscours
4 UT3 - Université Toulouse III - Paul Sabatier
5 LIMSI - Laboratoire d'Informatique pour la Mécanique et les Sciences de l'Ingénieur

The 2018 CLEF eRisk is composed of two tasks: (1) early de-tection of signs of depression and (2) early detection of signs of anorexia.In this paper, we present the methods we developed when participatingto these two tasks. We used two types of representations of the texts:one uses linguistic features and the other uses text vectorization. Theserepresentations are combined in different ways in models that are trainedusing a machine learning approach. These models are then used to builtthe 5 runs we submitted for task (1) and the 2 runs for task (2), whichdifferences are also detailed in this paper. For task (1), best results wereobtained when combining the methods based on features and text vec-torization, and for task (2), the method based on text vectorization givesthe best results.

Mots clés: Social media; Depression detection; Anorexia detection; Information retrieval; Natural language processing; Machine learning

Domaines: Apprentissage [cs.LG]

Faneva Ramiandrisoa, Josiane Mothe, Farah Benamara, Véronique Moriceau. IRIT at e-Risk 2018. 9th Conference and Labs of the Evaluation Forum, Living Labs (CLEF 2018), Sep 2018, Avignon, France. pp.1-12. ⟨hal-02290007⟩