Optimizing Social Media Sentiment Analysis: RoBERTa Framework for Evaluating Public Perception of DeepSeek

In the past few years, Artificial Intelligence (AI) has become a widely discussed topic by people. AI has affected various aspects of human life. DeepSeek is one of the AIs that has received a lot of attention. This study introduces an optimized framework for social media sentiment analysis to find out people's views of DeepSeek using Tweet data from X (formerly Twitter) taken in January 2025. Using a RoBERTa-based model, fine-tuned for social media text, and featuring an optimized preprocessing stage. One important result is that including stopwords during preprocessing significantly boosts sentiment classification accuracy to 87.5%. The sentiment analysis results show that people's opinions are very diverse and polarized, with a sentiment distribution of 41.1% being Positive, 36.8% being Neutral, and 22.1% being Negative. Word cloud analysis shows different views on DeepSeek. Some people appreciate DeepSeek's open-source accessibility while others express concerns about security and trust. These insights show that people's views on AI are complicated, making it much more vital to be transparent in AI development, set performance benchmarks, and build trust in AI.