Improving Sentiment Analysis in Social Media Using Deep Learning Techniques: A Comparison Study of Generative AI Model Approaches in Gujarati Language Context

Abstract

Social media's explosive growth has drastically changed how people express their thoughts,  feelings, and sentiments online. Low-resource languages like Gujarati are still not well studied, despite the significant advancements in sentiment analysis in widely spoken languages. In order to close this gap, this  study compares generative AI and deep learning methods for sentiment analysis in the context of Gujarati Language. A dataset that captured a variety of linguistic nuances, colloquial expressions, and code-mixed content typical of user-generated posts was assembled from multiple social media platforms. To ascertain of their respective advantages and disadvantages, a number of models, including LSTM, BERT, RNN, CNN, GPT-2, Gemma, LLaMA, RoBERTa, and Word2Vec-based architectures, were put into practice and method assessed. We used cross validation and standard evaluation metrics to test how well the model  performed.  The results showed that transformer-based and generative models work better than traditional sequence and the embedding-based models when handling complex context and meaning. This study adds new data and shows that generative AI can make models more robust and easier to interpret, which helps advance sentiment analysis for supervised represented languages. Our findings show strong results for Gujarati and open the door for future research on sentiment analysis in other regional and low- resource languages used on social media.