Automatic Segmentation and Summarization of Spoken Lectures

The ever-increasing number of online lectures has created an unprecedented opportunity for distance learning. Most online lectures are presented as unstructured text, audio and/or video files which make it di cult for students to locate relevant lectures and browse through them. In this thesis, we investigated several automatic lecture segmentation and summarization algorithms. Automatic lecture segmentation algorithms impose structure by inserting paragraph separators and section titles to the lecture to make them more readable. The Segmentation algorithms include K-Means, HMM, Hierarchical Segmentation. Unlike segmentation, Automatic lecture summarization algorithms compress lectures by selecting salient information based on importance of words/phrases and provide concise, coherent and readable summaries. The summarization algorithms were developed based on two supervised machine learning approaches: Support Vector Machine (SVM) and Conditional Random Fields (CRF). These approaches show comparable results with existing summarization methods. In this thesis we propose a novel Rhetorical Structure Index(RSI) to measure the structural importance of a feature. Experiments show that using RSI has significantly improved the segmentation accuracy as compared to the traditional contentbased feature weighting scheme such as TF/IDF. We also build summarization systems with structure-dependent models such that summarization generated for“Introduction” is di↵erent from that for “Main Lecture Content” and “Conclusion”. From all three segmentation algorithms HMM showed the best performance. However, similar to K-Means, HMM has definite number of segments which is not the case for Hierarchical Segmentation. Experiments show that using structure-dependent summarization outperforms the uniform summarization method by 15%.