Semantic Segmentation of MOOC Lecture Videos by Analyzing Concept Change in Domain Knowledge Graph

Long lecture video metadata needs to have topic wise annotation information for quick topic searching and video browsing. In this work we perform topical segmentation of long MOOC lecture videos to obtain start-time and end-time of different topics taught by the instructor. During teaching instructor uses different concepts to explain a topic. So instructor has his own way of selecting and binding these concepts to represent a topic. Additionally knowledge graph of a subject domain contains inherent domain knowledge. In this work we analyze how the instructor changes concepts during topic change, the inherent knowledge available in a domain knowledge graph, semantic similarity and contextual relationship between different concepts to perform topical segmentation of long lecture videos. As output, we get semantically coherent topics taught by the instructor along with their interval (start-time and end-time). We tested our approach on 61 long NPTEL [1] videos delivered on software engineering domain. Experimentally we find that the topic intervals generated by our system has \(\sim \)83% similarity with the intervals present in the ground truth. Holistic evaluation shows that our approach performs better than the other approaches in the literature.