Cross-lingual sentiment analysis with BERT

This thesis focuses on the use of the BERT model in sentiment analysis, a task that encompasses a variety of related topics. As such, this work provides an overview of some of the most important associated topics and classical approaches in this field. It begins with a brief introduction to the task of sentiment analysis and its subtasks, including opinion source identification, followed by a discussion of the most common approaches in this domain. A review of related works follows, including an explanation of the most frequently used techniques in the field of opinion mining. The subsequent sections delve into the details of the implementation and the datasets and models employed in this thesis. This thesis rigorously evaluated the BERT model's efficacy in sentiment analysis, particularly in multilingual settings, while comparing it against classical machine learning models using English and Farsi datasets. Findings revealed BERT's notable performance even with translated data, suggesting its potential in multilingual sentiment tasks where labeled data may be limited.