Comparative Analysis of Traditional and Large Language Models for Sentiment Analysis in the Serbian Language

Sentiment analysis is a field of natural language processing whose goal is to determine a person’s opinion about a product, movie, book, event, etc., based on a comment written in natural language. In this paper, we performed a comparative analysis of traditional and large language models (LLMs) for sentiment analysis in the Serbian language when it is conducted using text classification. The experiments showed the challenge we face when dealing with sentiment analysis in a language like Serbian, which is a complex and low-resourced language. Experiments have also shown that traditional approaches have an advantage over the used open-source LLMs, such as BERT or GPT-2. Although there is still a lot of room in this field for research and experimentation, The Davinci model from the GPT-3 generation has reached a very good accuracy of 91 percent.