COHETS: A highlight extraction method using textual streams of streaming videos

Abstract

As more and more conversation-oriented streaming videos become available, streaming platforms have gradually taken the place of traditional media for people to access information. Still, conversation-oriented streaming videos are often lengthy and people are reluctant to view the whole video. In this research, we investigated highlight extraction on conversation-oriented streaming videos. Previous highlight extraction methods analyzed visual features of videos and are therefore unable to deal with conservation-oriented streaming videos whose highlights are related to streamer discourses and viewer responses. For this reason, the proposed highlight extraction method called COHETS does not evaluate visual features but rather simultaneously examines textual streams of streamer discourses and viewer messages to extract meaningful highlights. Experiments based on real world streaming data demonstrate that streamer discourses and viewer responses via their feedback messages are useful for extracting highlights of conversation-oriented streaming videos. Also, our designed position enrichment and message attention techniques effectively distill the embeddings of the two textual streams and lead to extraction results that are superior to those of state-of-the-art deep learning-based highlight extraction methods and extraction-based text summarization methods.