Comparative analysis of methods for creating a sentiment lexicon of the Serbian WordNet

Purpose This study aims to present the adaptation and improvement of resources for sentiment analysis based on the Serbian WordNet (SrpWN), aiming to create more efficient sentiment lexicons for Serbian. Design/methodology/approach This research builds on the initial construction, analysis and replication of SentiWordNet: gloss vectorisation using a bag-of-words approach and classification using support vector machines (SVMs) and naive Bayes as baseline methods. Refinement was followed by replacing AdaBoost with SVM and applying state-of-the-art methods, such as recurrent neural networks, transformers and fine-tuned large language models. Findings The results showed that replacing traditional methods with a novel approach improved the accuracy of sentiment scores in SrpWN. The lexicons created using this method more accurately reflected the sentiment of word senses in Serbian. Originality/value This study advances sentiment analysis by enhancing and adapting existing methods for sentiment enrichment of SrpWN. The development of software tools and the construction of multiple sentiment lexicons represent significant progress in integrating sentiment information into WordNets, providing valuable resources for further research and applications in Serbian sentiment analysis.