Date depot: 12 avril 2023 Titre: Combining logical rules and statistical learning for opinion mining Directrice de thèse: Salima BENBERNOU (LIPADE) Encadrant : Mourad OUZIRI (LIPADE) Domaine scientifique: Sciences pour l'ingénieur Thématique CNRS : Données et connaissances Resumé: With the growing popularity of social media, the internet is replete with sentiment-rich data including reviews, comments, and ratings. Therefore, organizations are showing vivid interest in adopting sentiment analysis tools to exploit this data for decision making. For instance, it is used in politics for predicting election results, in marketing to measure user's satisfaction and in health care to predict the effect of drugs on patients. A sentiment analysis tool automates the process of extracting sentiments from a massive volume of data by identifying an opinion and deriving its polarity, i.e., Positive, Negative, or Neutral. In the last decade, the topic of sentiment analysis has also flourished in the research community and is becoming increasingly popular. Nevertheless, sentiment analysis of social media data is still a challenging task due to the complexity and variety of natural language through which the same idea can be expressed and interpreted using different texts. The aim of the thesis is to improve the accuracy of the results of sentiment analysis tools by using a semantic based approach. However, despite the advances in research in this area, constructing powerful tools that ensure end-to-end reasoning is still challenging since all learning algorithms are supervised, which is not the case for all systems. Even ChaTGPT improved some results but it is still challenging to have better accuracy, it is a data driven based approach (learning). The thesis aims at using the Statistical Relational Learning methods (rule based) which constitute a robust framework that associates relational representation and probabilistic learning. Moreover, since review documents in most cases contain several aspects and are related to many entities in the review, studying the inconsistency resolution mechanism in sentiment analysis to deal with documents with different aspects is the auspicious and challenging work direction.