JKU Thesis, Project and Seminar

old, I changed the topic later and the institue as well (IML)

Transcript

1. AI Master’s Thesis, Practicum and Seminar Multimodal Educational Video Summarization RAG Mahdi Khashan
2. Motivations - Students use multiple educational contents for learning like youtube playlists - Those playlists are lengthy to watch - There are duplication in them - Having context around a topic enhances learning - Student populate knowledge bases from multiple sources - Knowledge-Graphs in MultiModal RAG is not well explored - Agent-based RAG, explore interactive self-guided feedback
3. Applications - Q&A - Summarize Playlists - Concept Knowledge Graph
4. Methodologies Data Ingestion 1. Retrieve video + transcript + thumbnails/slides. 2. Extract audio embeddings for lecture style emphasis. Multi-Modal Encoding 3. Text encoder for transcripts and video titles. 4. Visual encoder for slides and key frames. 5. Audio encoder for emphasis (intonation/important points).
5. Methodologies Retriever (RAG) 1. Build a course knowledge base with per-lecture embeddings. 2. Enable retrieval at the video, section, or concept level. Generator (LLM) 3. Summarize at different granularities: ◦ Lecture-level summary ◦ Module-level overview ◦ Entire playlist synthesis 4. Explore a knowledge graph of concepts.
6. Challenges - Evaluations - Classroom videos, How to learn whiteboard contents - Reasoning (coherence and contextual understanding)
7. References - A Comprehensive Survey on Multimodal Retrieval-Augmented Generation (https://aclanthology.org/2025.findings-acl.861.pdf)