Enhancing Sentiment Analysis of Persian Tweets: A Transformer-Based Approach

In today's modern world, the prominence of social media is undeniable, with Twitter standing out as a pivotal platform for global communication and information sharing, particularly in expressing and amplifying emotions. In this digital era, sentiment analysis has emerged as a crucial tool for measuring emotions and reactions to daily events, especially in the context of business improvement. It enables businesses to rapidly and effectively decipher trends and customer insights. Within the realm of sentiment analysis, we encounter a diverse range of models, each with its unique features and limitations. This research aims to amalgamate the strengths of various approaches by integrating the Naive Bayes classifier, a bespoke rule-based model, and BERT—a relatively lightweight transformer model—particularly in the context of sentiment analysis of Persian Twitter media. Our findings reveal that traditional models such as SVM, Naïve Bayes, and MLP alone do not yield high-quality results. Our hybrid model, when used independently, outperforms BERT, achieving an accuracy of 89% compared to BERT's 86% which represents a significant advancement in sentiment analysis. Although slightly more structurally complex, it maintains computational intensity on par with BERT fine-tuning while outperforming BERT when used individually. This advancement stems from our unique approach of integrating Naive Bayes and a bespoke rule-based model, subsequently leveraging BERT for sentiment classification, thus enhancing its effectiveness in social media contexts.