王者荣耀比赛直播弹幕的情感分析研究：基于NLP技术的实证探索

NLP-Based Sentiment Analysis of Bullet Comments on Honor of Kings Live Streaming Matches

摘要: 随着游戏直播产业的蓬勃发展，分析海量弹幕中的用户情感对于理解观赛体验、优化内容运营具有重要的现实意义。本研究以热门MOBA游戏王者荣耀的职业比赛直播弹幕为对象，旨在准确识别并分析观众的情感倾向与讨论焦点。研究首先从Bilibili平台采集了20,704条弹幕数据，经过数据清洗、中文分词和停用词过滤等预处理后，采用TF-IDF方法进行文本特征提取。在此基础上，研究构建了支持向量机(SVM)分类模型进行情感分类，并与逻辑回归基线模型进行了对比。研究的核心内容在于系统性地评估了模型性能、分析了整体情感分布并挖掘了高频词汇背后的情感动因。实验结果表明，SVM模型在测试集上取得了0.8783的AUC值与0.7016的KS值，显示出优越的情感分类性能。情感倾向分析揭示，积极情感弹幕占比最高(64.49%)，表明观众对赛事内容普遍持正面态度。进一步的高频词汇与词云分析发现，观众的情感表达高度聚焦于“冠军”、“荣耀”等体现团队荣誉与游戏精神的词汇。本研究证实了自然语言处理技术在游戏弹幕情感分析中的有效性，不仅为相关方提供了洞察观众情感的量化工具，也揭示了弹幕文本中所蕴含的丰富集体情感与价值认同。未来研究可进一步结合比赛情境与多模态数据，以深化对用户情感动态的理解。

Abstract: The growing popularity of bullet comments in game live streaming has made sentiment analysis of massive comment data highly valuable for understanding viewer experience and optimizing content strategy. This study examines bullet comments from professional matches of the popular MOBA game Honor of Kings, aiming to accurately identify and analyze audience sentiment tendencies and discussion foci. To achieve this, 20,704 comment entries were collected from the Bilibili platform. After preprocessing steps including data cleaning, Chinese word segmentation, and stop-word filtering, Term Frequency-Inverse Document Frequency (TF-IDF) for text feature extraction was employed. A Support Vector Machine (SVM) classification model was subsequently developed for sentiment analysis and compared against a logistic regression baseline. The research systematically evaluated model performance, analyzed overall sentiment distribution, and explored emotional motivations behind high-frequency vocabulary. Experimental results showed that the SVM model achieved an Area Under the Curve (AUC) of 0.8783 and a Kolmogorov-Smirnov (KS) statistic of 0.7016 on the test set, demonstrating good sentiment classification performance. Sentiment analysis revealed that positive comments constituted the largest proportion (64.49%), indicating a generally favorable audience attitude toward the tournament content. Further analysis of high-frequency words and word clouds identified strong viewer focus on terms like “champion” and “glory”, reflecting team honor and gaming spirit. This study confirms the effectiveness of natural language processing techniques in game comment sentiment analysis, providing stakeholders with quantitative tools to understand audience engagement while revealing rich collective emotions and value identification embedded in data texts. However, future research could incorporate match context and multimodal data to deepen understanding of dynamic viewer emotions.