Constructing knowledge representation from lecture videos through multimodal analysis

E-learning has presented new opportunities for learning with the rapid development of information and communication technologies (ICTs). Learners are no longer restricted by the location and time to learn. Lecture video is one of the most common learning materials on e-learning platforms. It presents knowledge in a lively manner and keeps the learners more focused during the learning process. While organizing lecture videos in a sequential list seems to be a natural choice, it presents the problems of inefficiency in searching for domain concepts and the inability to show relationships between such concepts. In this work, the task of constructing a knowledge representation scheme for video corpuses is explored. The knowledge representation aims to achieve the goals of facilitating the searching of domain concepts and to extract the relationships between the concepts so as to identify effective viewing strategies for the corpus. A framework using text recognition, speech recognition, multimodal analysis and clustering techniques is proposed for the construction of the knowledge representation. Two lecture video corpuses on the topics of general chemistry and geometry are acquired from the Khan Academy for illustrating the feasibility of the proposed framework. Experimental results have shown that the framework can be used to achieve the goals under specific assumptions