Univrsity Of Tehran Press Journal of Information Technology Management 2980-7972 18 3 2026 07 01 Advancing Automatic Sarcasm Detection in the French Language 202 227 108011 10.22059/jitm.2026.414898.4482 EN Yannick U. Tchantchou Samen Assistant Prof., Senior Lecturer, Ph.D., Department of Mathematics and Computer Science, Faculty of Science, University of Maroua, P.O Box:814 Maroua, Cameroon; Laboratoire de Recherche en Sciences Informatiques et Applications (LRSIA), UAC, Abomey-Calavi, Benin. 0000-0002-4031-4317 Abou A. Djafar Ph.D. Candidate, Department of Mathematics and computer Science, Faculty of Science, University of Maroua, Maroua, Cameroon. Kaladzavi Guidedi Associate Prof., Department of Computer Science and telecommunications, National Advanced School of Engineering of Maroua, University of Maroua, Maroua, Cameroon. Fréjus A. A. Laleye Opscidia, 9 Rue des colonnes, Paris, France. Journal Article 2026 05 22 The proliferation of user-generated content necessitates robust sentiment analysis and opinion mining. Sarcasm detection presents a significant challenge in natural language processing, as sarcastic statements often invert literal meaning. Accurate sarcasm detection is crucial for enhancing sentiment analysis, opinion mining, recommender systems, and public opinion monitoring. While transformer-based and deep learning models have advanced sarcasm detection in English and other languages, French remains underdeveloped due to limited annotated datasets and the absence of models tailored to its linguistic nuances. Consequently, existing methods struggle to generalize and fail to leverage complementary semantic information to improve sarcasm recognition. This paper introduces a novel hybrid multi-task learning framework for French sarcasm detection. This architecture integrates transfer learning with sentiment analysis by combining a fine-tuned CamemBERT encoder, an auxiliary sentiment-analysis module, and a Bi-LSTM classifier. This approach jointly models contextual and affective information to better distinguish between sarcastic and non-sarcastic text. A significant contribution of this study is the development and release of a large-scale French sarcasm dataset comprising approximately 22,000 manually collected and curated annotated instances from diverse sources, providing a valuable benchmark for future research. Another key contribution is the proposed novel hybrid multi-task architecture that integrates transfer learning, sentiment analysis, and Bi-LSTM-based classification for improved sarcasm detection. Empirical evidence shows that sentiment-aware representations significantly improve sarcasm detection over standard transformer approaches. Experiments on the TransCasm dataset and the new corpus demonstrate the model's effectiveness, achieving F1-scores of 96% and 90%, respectively, surpassing prior results. https://jitm.ut.ac.ir/article_108011_cb19319caeb1084ebe48cc34834233ff.pdf