Generating YouTube Video Titles Using Closed Captions and BART Summarization
H.U. Senevirathne, Banage T. G. S. Kumara, Banujan Kuhaneswaran · 2024
YouTube is a popular platform for information sharing and entertainment. With billions of hours of video content uploaded and viewed daily, effective content organisation and discovery are critical. BART is a transformer-based model with unique properties that are ideal for summarising textual data. This study investigates BART’s ability to summarise textual content by incorporating video Closed Captions (CC) and using the popular natural language processing library Hugging Face to generate relevant and engaging titles for YouTube videos. The methodology consists of several steps designed to maximise the potential of BART summarisation. Initially, a diverse dataset of YouTube CC was gathered, representing a wide range of genres and content types. This dataset served as the foundation for training the BART Large CNN model. Hugging Face’s transformers library was then used to fine-tune the pre-trained BART model on a specific CC dataset, making it more adaptable to different YouTube content. The fine-tuning process involved adjusting the model’s parameters to match the properties of YouTube CC. It ensured that the model accurately captured the essence of the spoken content while remaining contextual. The trained and fine-tuned BART model effectively captures the essence of the video in concise and engaging headlines. For example, "my channel is back" is the generated title for the video with the title "I Was Hacked. But Now I’m BACK!" and "amazing facts about water soft vs. hard" is the generated title for the video with the title "Hard vs. Soft Water: What’s The Difference?". Cosine similarity was used to evaluate generated titles. The study has implications for content creators, viewers, and the entire YouTube platform, as it can improve content discoverability, increase user engagement, and simplify the content creation process.