| Abstract [eng] |
The sentiment analysis task has been widely studied and researched in Natural Language Processing. Various companies and resellers on the internet are becoming more interested in the latest automated sentiment analysis solutions. However, this task remains challenging, particularly in languages like Lithuanian, which are complex and subjective regarding sentiment expression. In this paper, we delve into the challenges of sentiment analysis in the Lithuanian language, particularly in the context of a relatively small dataset and a language that has been limitedly studied. After a comprehensive review of existing research in Lithuanian NLP tasks, we found that classical machine learning approaches and classification algorithms have limited capabilities when analyzing and detecting sentiments in the Lithuanian language on limited resources. This paper describes and analyzes the performance of pre-trained multilingual Large Language Models in Lithuanian sentiment analysis. This work explores sentiment analysis, its necessity, and its application possibilities. It thoroughly describes methods used in practice and research for sentiment analysis, including their possibilities and limitations. In this research, BERT and T5 models have been adapted for sentiment analysis of Lithuanian language documents. The paper concludes by analyzing and comparing the models' results, highlighting their effectiveness and practical implications for sentiment analysis in the Lithuanian language. |
