Web Implementation of Sentiment Analysis for Bulgarian Texts: A Case Study across Domains

This paper introduces a novel web application for sentiment analysis of user reviews written in Bulgarian, leveraging an innovative ensemble model combining Support Vector Machines (SVM) and Recurrent Neural Networks (RNN), created previously by the author. The system integrates a lexicon-based Random Forest (RF) classifier and an attention mechanism to optimize predictive performance. Experiments across three datasets—restaurants, hotels, and entertainment—demonstrate the model’s robust performance, achieving high accuracy, precision, recall, and $\mathbf{F 1}$ scores, with results consistently above $90 \%$. The focus of the paper is the analysis of the results, which highlights the importance of domain-specific vocabulary and corpus characteristics on model effectiveness. The web application offers real-time analysis for individual and batch reviews, showcasing its practicality for diverse applications. Future developments aim to enhance the lexicon, integrate real-time scraping, and improve the user interface, positioning the tool as a valuable resource for Bulgarian sentiment analysis research and applications.