Sentiment analysis of cultural differences in online comments on popular news

Abstract The rapid growth of online communication through social networks has created new opportunities for understanding public opinion on socially relevant issues. This research examines how sentiment analysis (SA) can reveal cultural differences, specifically analyzing Czech and Ukrainian online comments on news topics including the war in Ukraine, political discussions, public health issues (tick-borne diseases, COVID-19), LGBTQ+ community matters, and natural disasters. By comparing three large language models (GPT-3.5-Turbo, Twitter-XLM-Ro-BERTa, and Zephyr 7B) with native speaker evaluations, we assess whether AI-based sentiment analysis can accurately capture culturally-specific emotional expressions in medium-resource languages. Our dataset comprises 6,085 comments (2,999 Czech from X/Twitter, 3,086 Ukrainian from Telegram) collected during 2023, focusing on socially relevant news coverage. We employed a hybrid methodology combining machine learning analysis with expert validation by native speakers. The study addresses a critical gap in cross-cultural sentiment analysis research, as no previous studies have compared Czech and Ukrainian linguistic patterns in this context. Results demonstrate significant performance differences among models depending on language: GPT-3.5-Turbo achieved highest accuracy for Czech (p<0.001), while all models performed comparably for Ukrainian. Both populations showed predominantly negative sentiment (Czech: 69.93%, Ukrainian: 68.93% via GPT-3.5), reflecting shared emotional responses to crisis events.