Stories That Big Danmaku Data Can Tell as a New Media

Qingchun Bai, Qinmin Vivian Hu, Linhui Ge, Liang HeIEEE Access 2019 · OpenReview

Abstract

In online video communities, an emerging source of comment, such as danmaku, allows viewers to interact when watching. Prior work discussed the feasibility of application using danmaku, while the comprehensive analysis of large-scale data is vacant to be filled in. We here release our danmaku data collection and report interesting observed phenomena in the danmaku comments. This analysis reveals the temporal distributions and user's access patterns of online time-sync comments. In particular, we distribute two novel natural language processing (NLP) tasks based on our danmaku dataset and provide baseline models. In the first task, we show how the naive models predict positive or negative sentiment given a danmaku comment, which effectively extends the real applications such as opinion poll prediction and marketing investigation. In the second task, we propose to use the NLP summarization model to make video tagging and summarization. The experimental results suggest that danmaku can not only support deeper and richer interactions between viewers and videos but also with high research value.

Introduction: The Rise of Danmaku

Traditional comment sections are static and decoupled from the playback timeline. In contrast, danmaku—a medium originated in Japan's Niconico and popularized by Chinese platforms like Bilibili and Youku—allows viewers to overlay comments directly onto specific moments of a video. These "bullet comments" create a unique, time-synced social experience, yet research has historically lacked the large-scale datasets necessary to understand this phenomenon at scale. This paper introduces the DR dataset, a collection of over 57 million entries, to explore how these short, anonymous bursts of text can inform video content analysis and viewer behavior.

The DR Dataset: A Large-Scale Resource

The researchers developed a custom crawler to aggregate data from Youku, capturing text alongside critical metadata such as "play time" (the specific second in the video the comment appears) and user IDs. Two primary datasets were released:

  • DR_Four: A large-scale set covering four categories (Cartoon, Entertainment, Movies, and Opera) from late 2017.
  • DR_E (Extension): The flagship dataset containing 57,176,457 danmaku comments and 2,170,033 traditional comments from over 6.2 million users.

Unlike previous datasets, DR_E includes manual emotional polarity tagging and video synopses, providing a robust benchmark for NLP tasks.

The Nature of Danmaku Data

Analysis of the DR dataset revealed several unique characteristics of danmaku communication. One of the most striking is the "4-Character Rule": the average danmaku is incredibly brief, averaging just four Chinese characters. This brevity is driven by platform constraints and the desire for improvisational expression that does not obstruct the video content.

Visual representation of danmaku comments synced to specific video timestamps, illustrating the temporal resonance of viewer reactions.

Furthermore, the data shows Temporal Resonance. Danmaku frequency is not random; it peaks during moments of high emotional impact. In scripted content like films, these peaks align with major plot twists, whereas in talk shows, comments are more evenly distributed as users react to ongoing dialogue. Despite the reputation of bullet chats being "noisy," the study found that 83% of the data is meaningful text, proving its utility for data mining.

Natural Language Processing Tasks

Sentiment Analysis

To handle the unique noise of danmaku (slang, emoticons), the researchers applied various models to predict emotional polarity. The LSTM (Long Short-Term Memory) model achieved the best performance with an F-measure of 0.7105, significantly outperforming traditional Support Vector Machines (SVM) and Logistic Regression (LR). The results indicate that specialized emotion dictionaries are essential for parsing the informal language of danmaku.

Video Summarization and Tagging

The researchers treated video tagging as a text summarization task. They found that a Seq2Seq + Comments hybrid model—which utilizes both danmaku and traditional long-form comments—performed best, reaching a ROUGE-1 score of 0.1825. This suggests a synergy between the two media: danmaku provides the "pulse" of immediate reaction, while traditional comments provide the necessary descriptive context.

Conclusion and Future Work

While the brevity of danmaku presents challenges for deep semantic extraction, its high research value is undeniable. The authors propose that these NLP models could eventually power "real-time guidance" systems—automated prompts that suggest constructive topics at appropriate times to foster healthier community interactions and mitigate the "herding effect" of negative sentiment. By leveraging the temporal nature of danmaku, we can gain deeper insights into how audiences collectively experience digital media.

Summary updated

Abstract

In online video communities, an emerging source of comment, such as danmaku, allows viewers to interact when watching. Prior work discussed the feasibility of application using danmaku, while the comprehensive analysis of large-scale data is vacant to be filled in. We here release our danmaku data collection and report interesting observed phenomena in the danmaku comments. This analysis reveals the temporal distributions and user's access patterns of online time-sync comments. In particular, we distribute two novel natural language processing (NLP) tasks based on our danmaku dataset and provide baseline models. In the first task, we show how the naive models predict positive or negative sentiment given a danmaku comment, which effectively extends the real applications such as opinion poll prediction and marketing investigation. In the second task, we propose to use the NLP summarization model to make video tagging and summarization. The experimental results suggest that danmaku can not only support deeper and richer interactions between viewers and videos but also with high research value.
BibTeX
@inproceedings{bai2019stories,
  author = {Qingchun Bai and Qinmin Vivian Hu and Linhui Ge and Liang He},
  title = {{Stories That Big Danmaku Data Can Tell as a New Media}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=S6fI23WZ3bH},
}
Figure