Décrire les textes politiques par le deep learning : à la recherche de nouveaux observables
Résumé
The methods of deep learning, particularly deconvolution, have recently made it possible to go beyond simple classification tasks in order to develop text description tasks. This article proposes to apply the methods of deep learning on a corpus of electoral proclamations by the left and the right between 1958 and 2017. We highlight some salient results to interpret electoral speeches under the French V° Republic in order to emphasize, from a methodological point of view, the added value of deconvolution protocols. We thus point to well-known units of textual statistics (such as specificities, lexical and grammatical co-occurrences). But we also show how the model allows for the formal capture of complex syntagmatic units, which the analysis of textual data has often conceptualized without being able to identify them automatically, such as patterns (« motifs ») or passages.
Origine | Fichiers produits par l'(les) auteur(s) |
---|