Instructional Video Summarization with Transformers: A Curriculum Learning Approach for ASR-Generated Transcripts

Abstract

This paper addresses the challenge of abstractive summarization for instructional video transcripts. Utilizing a document-level encoder rooted in transformer architectures, the proposed methodology enhances the fluency and generalizability of generated summaries across diverse video content. A unique dataset of over 5,000 extracted transcripts supports the training process, employing specific fine-tuning and order-preserving techniques. Assessments based on metrics such as Content F1 and human evaluations confirm that the synthesized narratives achieve quality comparable to human-authored text, providing concise and informative overviews for online educational platforms.