In the digital age, where the proliferation of text and audio content is redefining modes of communication, the analysis of multilingual media dynamics is becoming a fundamental scientific challenge. Human interactions, whether written or oral, now unfold in a complex linguistic tangle, incorporating informal, dialectal expressions and code-switching discourse. Faced with this linguistic fragmentation, it is imperative to design advanced approaches capable of structuring, interpreting and exploiting these massive flows of information.
This thesis is positioned at the crossroads of three major axes: discourse trend extraction, multilingual automatic transcription and sentiment analysis in informal linguistic scenarios. In response to these challenges, we propose a new approach to decoding thematic dynamics based on probabilistic models (LDA, HDP) coupled with sequential predictive models (ARIMA), in order to extract latent trends and model underlying temporal dynamics, while analyzing the evolution of societal concerns. Next, we introduce a hybrid system for multilingual automatic transcription and keyword detection in continuous speech, based on an architecture incorporating optimized versions of Whisper (FasterWhisper and WhisperX), combined with speech activity detection models (Silero VAD and Pyannote VAD), guaranteeing greater resilience to linguistic and contextual variations. Furthermore, we combine residual neural networks (ResNet-18, ResNet-152) with acoustic (MFCC) and contextual (M-BERT) embeddings to optimize the rapid identification of critical information in emergency situations.
Finally, the aim is to propose a strategy for analyzing sentiment in multilingual content, where dialectal and code-switching exchanges, characterized by semantic ambiguities, complicate interpretation and represent an unprecedented challenge. In response, a hybrid architecture combining deep bi-directional neural networks (Bi-LSTM) and an ensemble learning approach (AdaBoost-SVM), incorporating adaptive cost weighting, improves the accuracy of opinion classification, while mitigating biases linked to imbalances in the corpora explored.