Evaluation of Transformer-Based Approaches for Sentiment Analysis in Uzbek

This paper presents a fairly relevant comparative analysis of two transformer-based language models, namely, RoBERTa and T5. These models were prepared and retrained specifically for sentiment analysis tasks in Uzbek texts. Due to the limited language resources for the Uzbek language, the authors formed a custom text corpus, which included tens of thousands of sentences. It should be noted that the sources for this corpus were reliable resources, including Uzbek Wikipedia and literature. In addition, both models were sufficiently finetuned and optimized for efficient operation. Moreover, experimental results showed that the RoBERTa-based model achieved higher accuracy, recall, and F1 score compared to the T5-based model. This study highlights the potential of transformer-based approaches. In addition to this, it should be noted, that authors also included necessary information about Uzbek language to make readers fully understood about the current scientific work. Besides, authors also conducted quite deep research about alternative scientific works, which have common ideas or thematic tasks with the current study.