基於面向的手機音樂節奏遊戲評論情感分析

台灣手機遊戲市場成熟且龐大，根據2020年App Annie的調查數據顯示，台灣手機遊戲營收位於全球第7，每位付費玩家產生的營收排名為全球第3 名。手機遊戲市場的蓬勃發展意味著更多玩家對於遊戲資訊的需求，遊戲公司也需要更多玩家的反饋，此時玩家所撰寫線上手機遊戲評論便成為最直接的資訊來源，如何在海量評論中快速取得有用資訊儼然成為一項重要課題。每一款遊戲吸引玩家的要素都不同，有些玩家注重遊戲體驗、劇情，有些注重人物設計、遊戲美術、音樂，有些則注重營運和福利多寡，本研究將這些玩家注重的要素視為遊戲的面向，作為分類的依據。我們挑選了手機遊戲中玩家對各面向重視程度較均衡的一款音樂節奏遊戲 “BanG Dream! 少女樂團派對”，蒐集其Google Play商店玩家評論，並使用在相關研究成果表現良好的LSTM深度學習技術進行面向分類，並對所有評論及面向評論進行情感分析。本研究整體評論的情感分析模型之F1-score為0.906、其中面向分類模型之F1-score平均為0.685、各面向評論的情感分析模型之F1-score平均為0.807。本研究除了能夠分析出評論中含有的玩家情感，同時也能解決評論星等評價可信度不足的問題。透過本研究呈現之綜合分析，玩家能夠快速地從海量評論中找到自己關注的資訊，遊戲開發公司與代理公司也不用額外請玩家填寫問卷調查來獲取資訊，而是透過模型分析大量玩家反饋來掌握該公司之遊戲在目前市場上的優勢、劣勢、機會與威脅。

關鍵字

手機音樂節奏遊戲；遊戲評論；面向；情感分析； LSTM

並列摘要

Taiwan's mobile game market is mature and huge. According to App Annie's survey in 2020, Taiwan was ranked 7th in the world in terms of mobile game revenue and 3rd in terms of generated revenue per paid player. The booming mobile game market means more players demand for game information, and game companies need more feedback from players. Online mobile game reviews written by players are the most direct source of these information. How to quickly extract useful information from these huge number of reviews is a very important research topic. Some players focus on game experience and storyline, some on character design, game art and music, and others on game company’s operation and benefits. In this study, these are called aspects of mobile games. We select "BanG Dream! Girls' Band Party", a music rhythm game with balanced emphasis on these aspects, as target of our case study. We collect its player reviews from the Google Play Store, and used LSTM, a deep learning technique with good performance in related research, to perform aspect classification and sentiment analysis for the whole review and for aspect-classified review. The F1-score of our sentiment analysis for the whole review is 0.906, the average F1-score of our aspect classification is 0.685, and the average F1-score of our sentiment analysis for aspect-classified review is 0.807. Our study not only can analyze the players' sentiment contained in the reviews, but also can solve the problem of insufficient credibility of star rating for the reviews. Through our comprehensive analysis, players could quickly find out the information they care about from the huge number of reviews. In addition, game developers and agents could analyze a large number of players' feedback to grasp the strengths, weaknesses, opportunities and threats of the company's games in the current market, without asking players to fill out questionnaires.