文化产品网络评论的跨平台与多场景特征比较——以电影《哪吒之魔童闹海》为例

Cross-Platform and Multi-Scenario Comparison of Online Reviews for a Cultural Product—A Case Study of “Ne Zha 2”

摘要: 为比较文化产品网络评论在不同平台与讨论场景中的分布特征，本文以电影《哪吒之魔童闹海》为例，对微博、知乎和Bilibili样本进行三分类情感分布及高频词比较。研究使用两组用途不同的数据：跨平台样本共4657条，包括微博310条、知乎3178条和Bilibili 1169条；Bilibili场景样本来自三个主题视频，共4482条。随稿材料仅保留正向、中性和负向标签及汇总结果，未保留原分类模型、训练语料和验证记录，因此本文将标签作为既有编码开展描述性统计，不推断模型性能。结果显示，微博、知乎和Bilibili样本的中性评论占比分别为54.8%、63.4%和62.8%，正向评论占比分别为34.8%、28.0%和31.7%，负向评论占比分别为10.3%、8.6%和5.6%。在三个Bilibili视频样本中，“隐喻分析”视频的正向评论占81.7%，而“好不好看”和“电影解说”视频的负向评论分别占56.7%和61.7%。研究表明，平台语境、视频主题、受众选择与观察窗口可能共同影响样本中的情感分布；在样本非随机且标签生成过程不可复核的条件下，相关差异应解释为探索性的样本特征。

Abstract: This study compares the distributional characteristics of online reviews for a cultural product across platforms and discussion scenarios, using the film “Ne Zha 2” as a case. It examines three-class sentiment distributions and high-frequency terms in samples from Weibo, Zhihu, and Bilibili. Two datasets serve distinct purposes: a cross-platform sample of 4657 comments (310 from Weibo, 3178 from Zhihu, and 1169 from Bilibili) and a within-Bilibili sample of 4482 comments drawn from three topic-oriented videos. The available research materials retain only positive, neutral, and negative labels and aggregate results; the original classifier, training corpus, and validation records were not archived. The labels are therefore treated as pre-existing codes for descriptive statistics, and no model-performance claim is made. Neutral comments account for 54.8% of the Weibo sample, 63.4% of the Zhihu sample, and 62.8% of the Bilibili sample. The corresponding positive shares are 34.8%, 28.0%, and 31.7%, while the negative shares are 10.3%, 8.6%, and 5.6%. In the three-video comparison, positive comments account for 81.7% of the metaphor-analysis sample, whereas negative comments account for 56.7% of the “Is it good?” sample and 61.7% of the film-commentary sample. The results suggest that platform context, video topic, audience selection, and observation window may jointly shape the observed distributions. Given non-random sampling and an untraceable labeling pipeline, the differences should be interpreted as exploratory sample characteristics.