The explosive growth of educational video content has created both opportunities and challenges for educators and learners. While educational videos have become a cornerstone of modern education, particularly in asynchronous environments, the time-intensive nature of video editing and the cognitive load of processing lengthy content remain significant barriers. This thesis introduces MMVSkim, a novel multimodal video skimming architecture specifically designed for low-production educational environments, where critical information is often conveyed verbally rather than through sophisticated visual production. 

The MMVSkim architecture employs a dual-metric relevance assessment approach that integrates both textual and visual features. The system leverages advanced speech recognition for accurate transcript generation, deep learning-based text summarization as a reference point for sentence-level relevance evaluation, and shot detection for structural analysis. 

By assigning weighted importance scores to each sentence based on its semantic significance and visual context, the MMVSkim architecture identifies and preserves essential educational content while eliminating low-information segments such as filler speech, tangential discussions, and extended pauses. Empirical evaluation on the TVSum50 dataset, while not specifically suited for the educational domain, demonstrates the effectiveness of our approach, achieving an average AUC-ROC score of 0.6 for lecture-style videos. 

This performance, while modest in absolute terms, represents meaningful discrimination ability in educational contexts compared to random selection (0.5). Further analysis reveals that MMVSkim’s specialized design favours educational content, performing significantly better on lecture-style videos than on other video categories.

The primary contributions of this research include:

1. A lightweight multimodal architecture tailored specifically to educational video content

2. An adaptive weighting mechanism that balances textual and visual features according to content characteristics

3. A flexible deletion metric that allows for variable compression ratios while maintaining instructional coherence By providing both automation and customizable parameters, MMVSkim eliminates the time-consuming task of manual video editing while empowering educators to optimize the framework for diverse educational video styles. 

This contribution to educational technology allows instructors to focus on content creation and student engagement rather than post-production tasks, addressing a critical pain point in increasingly video-centric educational environments.