A Proposal of Sentiment Analysis Approach for Comments Based on Semi-supervised Learning and Topic Analysis

This study proposed an approach that can indi-vidually determine the corresponding sentiments for each topic mentioned in the comments. As a case study, we collected comments from 2 battle-royale games from the Steam platform. Using the results from this study to provide advice for game developers on areas where their games need further improvement and update, or develop new games via mining information from user comments of already released same genre games. We first fine-tuned the Bidirectional Encoder Representations from Transformers (BERT) model for classifying comments that are not related to games or do not contain enough content for thematic analysis. Second, utilizes Latent Dirichlet Allocation (LDA) to extract topics from comments. At last, employ another BERT model to analyze user sentiment toward every topic in every comment. The experimental results can be obtained by comparing topics of different games which were scored by sentiment analysis. The highest score in the game NARAKA is the "Optimization" topic with 0.4964. In the game Z1, the "muti-gameplay" topic is 0.4393.