Political discourse in the digital age takes many forms, including news articles, social media posts, political speeches, and public opinion polls. Understanding popular sentiments, analyzing political trends, and evaluating the emotional landscape surrounding political problems all need a careful examination of the attitudes conveyed in various textual sources. However, the intrinsic complexities of language, contextual complexities, and the nuanced interaction of multiple emotions provide severe hurdles to sentiment analysis in political works. Traditional sentiment analysis approaches frequently fail to capture these subtle sentiments properly. The core research question is concerned with the adaption and optimization of transformer models to provide precise sentiment analysis across a wide range of political text forms. Unlike many domains of NLP, the political arena contains nuances that are not easy to identify by machines, thus this study aims to use three transformer models, BERT, RoBERTa and GPT-3 to investigate which one of the models is best to encapsulate the complex scenario of the political arena based on the NewsMTSC dataset, which is curated from two distinct datasets, POLUSA and BiasFlipper. The results show that the best performing of these models is RoBERTa with an average accuracy of 84.97% and F1-score of 86.98%. While BERT scores are closer to RoBERTa with average accuracy and F1-score of 81.72% and 81.09% respectively, GPT-3 had the worst performance with an average and F1-score of 79.40% and 79.51%. These results show that with advancements in technology and machine learning, accurately classifying data such as in politics will not always give the best results due to the complexity of human emotions in such cases.
