LLM-Based Sentiment Analysis of RPG Genre Game Reviews on Steam

The rapid growth of the digital gaming industry has led to a significant increase in user-generated content, particularly reviews on platforms like Steam. Analyzing the sentiment of these reviews offers valuable insights into user experiences and provides developers with data to support decision making. This study examines the effectiveness of three prominent large language models (LLMs), BERT, RoBERTa, and DistilBERT, in classifying sentiment within RPG game reviews sourced from Steam. The models were evaluated across performance metrics including accuracy, precision, recall, and F1-score, under varying learning rates. Experimental results reveal that RoBERTa outperforms BERT and DistilBERT, achieving the highest F1-score of 97.50% at a learning rate of 2e-5. BERT follows with a peak F1-score of 95.50%, while DistilBERT achieves a competitive score of 95.30%. The impact of text preprocessing was also analyzed, showing that excessive transformations such as lemmatization and stopword removal may negatively affect performance. These findings provide important insights for optimizing sentiment analysis methods in the context of user reviews and contribute to advancing natural language processing (NLP) applications on digital platforms. The study also introduces a domain-specific sentiment lexicon analysis to reflect on the contextual ambiguity of common gaming terms, offering a valuable foundation for future research in sentiment analysis for online gaming communities.