Topic Modeling and Sentiment Analysis of YouTube Comments on OpenAI’s Stargate Project Using BERTopic and RoBERTa

Abstract

The rapid development of generative artificial intelligence has sparked growing interest, widespread adoption across various sectors, and intense public debate, particularly following the announcement of the Stargate project by OpenAI, SoftBank, and Oracle. The diverse public reactions documented on social media platforms, particularly in YouTube comment sections, can serve as a valuable source of public opinion data. The YouTube comment data used in this study was collected through scraping and subsequently processed through a cleaning procedure, yielding clean data covering the period from January 21, 2025, to April 21, 2026. This study aims to identify dominant discussion topics and the underlying sentiment to uncover indications of public concerns and expectations regarding the AI development project led by OpenAI through the Stargate initiative. The analysis in this study employs a structured NLP approach, specifically BERTopic for topic modeling and RoBERTa for sentiment analysis. The RoBERTa model was fine tuned for binary classification using a combination of pseudo labeling, annotation, class balancing, and classification decision threshold optimization, which optimally yielded an accuracy of 0.89 and a macro F1 score of 0.87 in the two class evaluation. BERTopic successfully identified 43 topics with a coherence score of 0.58 and a topic diversity of 0.71.