Abstract
This paper investigates sentiment analysis of tourist reviews about Kazakhstan using 11,125 English comments from YouTube travel channels. We compare four sentiment analysis approaches NLPTown’s BERT model, TextBlob, VADER, and Stanza, evaluating their accuracy and robustness across different linguistic contexts, short vs. long reviews, sarcasm, neutral tone. Each algorithm’s strengths and weaknesses are examined: lexicon-based tools (TextBlob, VADER) are fast and effective on short, informal texts, but struggle with subtle nuances like sarcasm, whereas advanced NLP models (Stanza, NLPTown) handle longer, context-rich reviews more effectively. Classification performance is assessed with standard metrics to ensure a fair comparison. In addition, we explore hybrid ensemble methods, majority voting, weighted voting, and stacking, to improve sentiment classification. Our findings show that ensemble models significantly outperform individual algorithmsijirss.com, particularly in addressing class imbalance and mixed sentiments (reviews containing both positive and negative cues). Notably, the best results are achieved by a stacking ensemble using a Random Forest metaclassifier, which achieved the highest overall accuracy and F1score in our experiments.