Automatic Chapter Generation for Hindi-English YouTube Videos

Abstract

Chaptering long-form multilingual videos into semantically meaningful chapters is crucial to making the content more accessible and navigable, especially in educational and informative videos. Previous work such as VidChapters-7M has investigated large-scale chaptering for English videos, but little research has ventured into bilingual or code-switched videos, especially for underrepresented languages such as Hindi. In this paper, we propose a new pipeline for automatic chapter generation on Hindi-English YouTube videos by utilizing both audio transcripts and semantic information. We construct a dataset of code-switched videos and assess our chapter generation using BLEU scores, readability (with Flesch Reading Ease), and temporal coherence metrics. We also include human evaluation to assess relevance and coherence of chapters. Our experiments show encouraging performance, and we highlight the challenges and opportunities of bilingual video understanding. We declare that this paper opens new avenues for multilingual video segmentation, paving the way to the development of inclusive AI systems and improved content navigation tools for the Indian subcontinent and beyond.