ED Mathématiques et Informatique
EXTRACTION, TRANSCRIPTION AND ANALYSIS OF MULTILINGUAL PUBLIC OPINIONS FROM SOCIAL NETWORKS
by Samawel JABALLI (LaBRI - Laboratoire Bordelais de Recherche en Informatique)
The defense will take place at 10h00 - Amphithéâtre 3 Université de Bordeaux, Bâtiment A9, 351 cours de la Libération, 33405 Talence Cedex
in front of the jury composed of
- Henri NICOLAS - Professeur des universités - Université de Bordeaux - Directeur de these
- Lilia CHENITI-BELCADHI - Professeure des universités - ISITCOM (Institut Supérieur d'Informatique et des Technologies de Communication) - Université de Sousse - Rapporteur
- Zied KECHAOU - Professeur des universités - Université de Kairouan - Examinateur
- Mounir ZRIGUI - Professeur des universités - Université de Monastir - Directeur de these
- Jenny BENOIS-PINEAU - Professeure des universités - Université de Bordeaux - Examinateur
- Ahmed BOUNEKKAR - Maître de conférences - Université Claude Bernard Lyon 1 - Rapporteur
In the digital era, where the proliferation of text-audio content is reshaping communication paradigms, the analysis of multilingual media dynamics stands as a major scientific challenge. Human interactions, whether written or spoken, now unfold within a complex linguistic network, incorporating informal expressions, dialectal variations, and code-switching discourse. Faced with this linguistic fragmentation, it is imperative to develop advanced approaches capable of structuring, interpreting, and leveraging these massive flows of information. This thesis lies at the intersection of three major research axes : the extraction of discursive trends, multilingual automatic transcription, and sentiment analysis in informal linguistic scenarios. In response to these challenges, we propose a novel approach to decode thematic dynamics by leveraging probabilistic models (LDA, HDP) coupled with sequential predictive models (ARIMA), aiming to extract latent trends and model underlying temporal dynamics, while analyzing the evolution of societal concerns. Subsequently, we introduce a hybrid system for multilingual automatic transcription and keyword spotting in continuous speech, relying on an architecture integrating optimized versions of Whisper (Faste rWhisper and WhisperX) coupled with voice activity detection models such as Silero VAD and Pyannote VAD, ensuring greater resilience to linguistic and contextual variations. Furthermore, we incorporate residual neural networks (ResNet-18, ResNet-152) with acoustic embeddings (MFCC) and contextual embeddings (M-BERT) to enhance the rapid identification of critical information in crisis situations. Finally, the objective is to develop a sentiment analysis strategy tailored to multilingual, dialectal, and code-switched content, where semantic ambiguities add complexity to interpretation, posing an unprecedented challenge. To address this, we propose a hybrid architecture combining bi-directional deep neural networks (Bi-LSTM) with an ensemble learning approach (AdaBoost-SVM), integrating cost-sensitive weighting to enhance opinion classification accuracy while mitigating biases associated with imbalanced corpora.