Learning to Summarize YouTube Videos with Transformers: A Multi-Task Approach

The wide-range availability of online video content has made people consume more time to keep up with the amount of information available. In this paper, we present a system for summarizing YouTube videos that use NLP and machine learning techniques to retain the important details without any data loss. There are a few techniques for summarization on YouTube, the first approach is YouTube API which has been a popular method used for video summarization. However, it has certain limitations, such as limited accuracy, language barriers, and inability to summarize caption-less videos. To overcome these limitations, a new approach using transformers has been proposed. The proposed approach involves downloading the video' audio, converting it to WAV format, performing speech-to-text conversion using the Hugging Face Automatic Speech Recognition model, and then using transformers and pipeline for summarization. For evaluation we have used Rogue (Recall-oriented understudy for Gisting evaluation) which is a popular metric used for summarization tasks with an f1 score of 0.6. The approach has been tested on multiple videos with different video lengths and has shown good results in terms of accuracy and efficiency. The proposed approach can be used for a wide range of applications, including online education, video marketing, and false thumbnail videos.