Analyzing Genshin Impact Review Scores Using BERT Model

Genshin Impact, a critically acclaimed ongoing multiplayer RPG, has amassed 4.4 million reviews on Google Play, reflecting its widespread popularity across mobile and PC platforms. Developers face the challenge of handling a growing amount of user feedback, which slows down the timely integration of valuable insights into the game. To address this, sentiment analysis is conducted on 48,000 English reviews using the BERT method, providing a comprehensive overview for developers. This study delves into hyperparameter adjustments for the BertForSequenceClassification model in predicting sentiment for Genshin Impact reviews. The best accuracy achieved by our fine-tuned model is 70 %, with a Matthews coefficient score of 55 %, utilizing the bert-base-uncased model (4 epochs, batch size 32, learning rate 0.00002). While successfully constructing a BERT model for sentiment prediction, the study recognizes room for improvement and recommends further fine-tuning and increasing the dataset size to exceed the accuracy obtained by this study.