- 學位論文
深度學習框架應用於Steam 遊戲評論分類
Application of Deep Learning for Steam Game Review Classification
摘要
遊戲評論中隱藏了許多有用的資訊，像是玩家對遊戲平衡機制的建議、遊戲運行時所遇到的錯誤情形，又或是被玩家讚譽有加的遊戲特色，對於遊戲開發商來說，這些評論資訊在遊戲開發、營運、維護上是極具價值的決策參考依據。然而，對部分獨立(Indie)遊戲廠商而言，並無多餘的人力、財力來對遊戲評論進行資訊檢索、分析，因此，本研究希望能利用深度學習與預訓練模型來幫助遊戲廠商進行自動化評論分類，快速地從繁多的遊戲評論中挑出包含特定類別資訊的評論，進而拓展遊戲評論應用的可能性。本研究以數位遊戲平台Steam上的英文遊戲評論作為模型訓練資料，並以多標籤標註的方式將評論標註上Suggestion、Pro、Con、Bug等標籤；在分類策略方面，則根據資料集特性、分類方式與分類模型等三個方向，規劃出多種分類策略。最終發現，使用RoBERTa預訓練模型與簡單的神經網路對多標籤資料集進行訓練，所建構的單一分類模型能得到達到最好的成效，Pro、Con與Bug等類別F1-Score皆高於0.8，而Suggestion類別也接近0.75，與過往研究比較也來得優異。因此，透過本研究的自動化評論分類，遊戲開發商可對遊戲評論進行快速且精準地分類，進而加速評論分析流程，來更好地發展商業應用。
並列摘要
Game reviews often contain valuable information, such as players' suggestions on game balance, bugs during gameplay, or game features that are highly popular. These reviews serve as invaluable references for game developers during game development, operation, or maintenance. Nevertheless, some Indie game developers may lack the personnel and resources to retrieve review information and conduct further analysis. Therefore, this study aims to use deep learning and pretrained models to assist game developers in automatic game review classification, to swiftly identify the reviews that contain specific categories of information from a large number of game reviews, and thereafter broaden the potential applications of game reviews. In this study, English game reviews on the digital game platform Steam were used as training data. The reviews were labeled with categories such as Suggestion, Pro, Con, and Bug via a multi-label annotation approach. And the classification strategies are designed based on dataset characteristics, classification methods, and classification models. Our findings indicate that using the RoBERTa pretrained model and a simple neural network to train the multi-labeled dataset resulted in the best performance after all. The F1-Scores for categories such as Pro, Con, and Bug were all above 0.8, and the Suggestion category was close to 0.75, which compares favorably to previous research.Therefore, through the automatic game review classification in this study, game developers can rapidly and precisely classify game reviews, accelerating the review analysis process to better develop commercial applications.