HYTAS: An AI-Powered Hybrid Extractive-Abstractive YouTube Transcript Summarizer

Abstract

Online lectures, tutorials, interviews and technical presentations are frequently available online, and such information can be more than people can process. The transcripts for YouTube videos are difficult to summarize due to their length, repetitive content, weak punctuation and potential for automatic speech recognition inaccuracies. The paper introduces HYTAS, a hybrid extractive-abstractive framework that starts by either retrieving captions or applying speech recognition to extract sentences, reconstructs coherent sentences, identifies topic boundaries, identifies the most relevant evidence using lexical and semantic features, filters out extraneous sentences via maximal marginal relevance, and generates structured summaries using a transformer model. A custom dataset of 180 videos was prepared from YouTube and split into 120 training, 20 validation, and 40 test videos. A dataset of 180 English-language YouTube videos was prepared and divided into 120 training videos, 20 validation videos and 40 test videos. HYTAS achieved ROUGE-1, ROUGE-2, and ROUGE-L F1 scores of 46.72, 23.88, and 43.15, respectively. It also achieved a BERTScore F1 of 0.906, a QAFactEval score of 0.887, and a topic coverage of 91.8%. The average compression ratio of the generated summaries was 8.6% while the redundancy rate was very low at 4.6%. The findings suggest that the use of source-grounded extraction and controlled abstractive summarization can yield concise, semantically coherent and factually consistent summaries that retain connections to the source video content.