Abstract
Time-Sync Comments (TSCs), commonly known as "danmu," have transformed the online video experience by allowing users to post comments synchronized to specific timestamps. However, this interactivity is often marred by spoilers that reveal future plot points. This paper, originally presented at CIKM '19, introduces SBN-IVA, a specialized deep learning framework designed to detect spoilers by leveraging the unique temporal and interactive properties of TSCs. By comparing a comment's similarity to its immediate neighbors versus its similarity to video "keyframes" (the climax), and employing a novel Interactive Variance Attention (IVA) mechanism to filter noise, the model significantly outperforms general-purpose text classification methods.
Introduction
Time-Sync Comments (TSCs) differ fundamentally from standard movie reviews or static comments. They are typically short, laden with network slang, and highly interactive, often forming conversational chains or "herding effects" where one comment triggers a series of similar reactions.
Existing spoiler detection methods, such as keyword filtering or Latent Dirichlet Allocation (LDA), often fail in this context because they do not account for the temporal flow of the video or the high volume of "noise"—emotional outbursts like "233333" (LOL) that lack semantic plot relevance. The core challenge lies in distinguishing between a legitimate reaction to the current scene and a spoiler that prematurely reveals information from the video's climax.
Methodology
The authors propose the Similarity-Based Network with Interactive Variance Attention (SBN-IVA). The model operates on the intuition that a spoiler is essentially a piece of "future" content appearing in the "present."
1. Word-Level Attentive Encoder
The process begins with a Bi-LSTM that utilizes an attention mechanism to weigh keywords (e.g., "murderer," "identity") more heavily than generic stop words or slang, creating a robust semantic representation of each comment.
2. Similarity-Based Network (SBN)
The SBN calculates two critical scores for a target comment:
- Neighbor Similarity: This measures how much a comment relates to the $R$ comments immediately preceding it. High similarity suggests the user is reacting to the current scene. A time decay function is applied here, as older comments lose relevance to the immediate interactive moment.
- Keyframe Similarity: The model automatically identifies "keyframes"—periods of high comment density in the final quarter of the video where resolutions typically occur. If a comment appearing early in the video matches the content of these keyframes, it is flagged as a potential spoiler.
3. Interactive Variance Attention (IVA)
To handle the inherent noise in TSC streams, the IVA mechanism calculates the similarity variance of neighboring comments. Noisy comments typically exhibit weak semantic relevance to their surroundings, resulting in high variance. The IVA assigns lower weights to these noisy neighbors, ensuring they do not skew the Neighbor Similarity score.
Experiments and Results
The researchers evaluated SBN-IVA using a massive real-world dataset crawled from Youku and Bilibili, spanning TV series, movies, and sports. The data was pre-processed to normalize slang and manually labeled to establish a ground-truth benchmark.
SBN-IVA outperformed all baselines, including keyword matching and the previous state-of-the-art deep learning model (DN-GAA), achieving an average 11.2% improvement in F1-score. The model performed exceptionally well on TV series (F1: 0.850), where plot-driven spoilers are most prevalent. Ablation studies confirmed that both the time decay function and the IVA mechanism are critical; removing either led to a significant drop in detection accuracy.
Conclusion and Future Work
SBN-IVA provides a robust solution for real-time spoiler detection by treating spoilers as statistical anomalies relative to their temporal context. While highly effective, the authors note limitations in cases of "spoiler attacks" (where many users post the same spoiler simultaneously). Future research may incorporate User IDs to track serial spoilers or integrate multimodal features (visual and audio cues) to more precisely identify keyframes and scene transitions.
Summary updated