Comparison of transformer-based models in sentiment analysis

Sentiment analysis is an artificial intelligence task that involves extracting emotions from texts. In recent years, the increasing interest and studies in this field have significantly contributed to the development of sentiment analysis research. Various machine learning algorithms, deep learning methods, and more recently, transformerbased models have become widely used in sentiment analysis tasks. This study was conducted to compare transformer-based models for use in sentiment analysis problems. Two separate datasets were used in the study: one consisting of 61121 entries and the other consisting of 5937 entries. The datasets used in the research were first subjected to cleaning and preprocessing steps to prepare them for use with transformer-based models. Subsequently, separate notebooks were run for ten different transformer models, and their results were documented. Each notebook was run for five epochs, and results were compared for both the large dataset and the smaller dataset with fewer samples. As a result of the study, the performance of transformer-based models was compared, and their outcomes under various parameters and metric usages were observed. In general, the high performance and fast processing capabilities of transformer models have demonstrated that they should be widely employed in sentiment analysis tasks. Thus, just as in many other Natural Language Processing (NLP) tasks, transformer-based models are expected to continue delivering successful results in sentiment analysis as well. With further parameter tuning and model-specific optimizations, highly performant models could be utilized in the future, achieving even greater success in real-world applications. Keywords: Natural language processing, Sentiment analysis, Transformer models