A Scene Matching of User-Generated Real-Time Commentsby Predicting Outgoing Time-Lag

Abstract

Knowing which scenes in a television program’s content have generated the greatest response is important for analyzing and improving program production. In particular, comments posted while watching an on-air television program (real-time comments) are suitable for analysis, as they are feedback posted immediately after viewing the footage. Previous research has focused mainly on either the content of the text or the time of posting, making it difficult to properly match long comments and other comments that take time to post to scenes. Therefore, we propose a novel method to match real-time comments to the scenes they refer to. We hypothesized that the time difference between the time a real-time comment is posted and the program scene it refers to, is an effective feature, and we set up a regression subtask to predict this feature as a continuous value and modeled it. We found that by utilizing this subtask, we can significantly improve the accuracy of scene-specific clustering for real-time comments. The proposed method is capable of utilizing both text and time of posting, and it was effective to adopt a neural network that integrates a word embedding layer based on contrastive learning and a fully connected recurrent layer as a model for such subtask. The results of this research will enable the evaluation of program scenes posted on SNS through comments from users who have a deeper understanding of the program’s content, which may be useful for improving video expression in a more diverse way than ever before.