AND MACHINE LEARNING
]]> Several models were tested, including Logistic Regression, Support Vector Machines, GRUs, and transformer-based models like RoBERTa and DeBERTa. The best performance came from RoBERTa with label smoothing and DeBERTa-v3, both reaching 98% accuracy. While these advanced models were the most accurate, simpler models like GRU were still competitive and easier to interpret.
The thesis also examined linguistic differences between real and AI-generated reviews. Real reviews were shorter, used more personal and emotional language, while AI-generated ones were longer, more structured, and often overused formal or generic phrases.
A working browser extension was built as part of the project. It allows users to analyze reviews directly on websites and see predictions with basic explanations. Although the tool works well, there are still limitations, such as handling newer AI models and providing clearer feedback for non-technical users.
]]>