Details

Optimizing Moment Retrieval for Educational Videos

Year: 2026

Term: Winter

Student Name: Julie Wechsler

Supervisor: Olga Baysal / Sriram Subramanian

Abstract: This thesis designs a moment retrieval model that is designed specifically for educational videos. Moment retrieval or temporal grounding refers to the task of retrieving the relevant segment in a video given a natural language query. There are currently no temporal grounding datasets with specifically educational videos, and because of this existing methods are tailored for short action based videos instead of more educational content. Throughout this thesis I create an educational dataset by selecting only educational videos from a much larger dataset, then create a temporal grounding method tailored specifically to educational videos. This approach first segments the video into topic segments, then uses the topic boundaries as candidates for a BiLSTM based temporal grounding model.