Smart Video Retrieval and Question Answering System
Likhith, Mahima Dubey, Mahim Dashora, V Kusuma Shree, R. Jayashree · 2023
More and more video content is being produced on daily basis and utilizing it to its fullest has never been easier. In this paper we propose a new solution for this problem as a smart video retrieval system, leveraging the progress made in extractive question-answering models like RoBERTa and DistilBERT. The system is designed mainly for educational videos in English, with major use cases in schools and colleges. The problem is divided mainly into two parts: 1. Video retrieval and 2. Timestamps retrieval within the video. The paper discusses different variations of the proposed architecture involving slide-text extraction, TFIDF and semantics-based indexing while also speaking about the implementation. We also introduce two new datasets based on educational YouTube videos and Cloud-computing course videos in the paper which are used to generate results and evaluate performance across all the variations implemented. We are able to achieve the best of 89.35% and 84.83% accuracy in video retrieval, and just 1% and 1.7% error in timestamp retrieval considering outliers across the two datasets respectively.