Enhancing Knowledge Graphs with LLMs: Keyword Extraction and Synonym Merging
summary
LLM pipeline extracts ~5,000 candidate keywords from laureate biographies, lectures, press releases and Wikipedia, then computes pairwise embeddings (≈25 million pairs) and filters with a cosine similarity threshold ~0.7 to surface synonym candidates; each pair is sent back to an LLM to merge by selecting a consolidated term, iterated twice to shift many singletons (≈2,900) into a target 3–10 laureates per keyword range, then loaded into Neo4j for visualization and expert review for museum use.
AI-written summary · about
chapters (8)
- 0:00Introduction and project overview
- 3:25Knowledge Graph enhancement process
- 6:14Keyword extraction with LLMs
- 8:41Synonym merging with embeddings
- 11:55Graph construction and evaluation
- 13:59Advantages and final results
- 14:44Example of enriched graph
- 15:07Q&A and resources
related talks
46mKnowledge Graphs + LLMs: What's the latest in research?
44mGive Your LLMs a Left Brain
23mKnowledge Graphs and LLMs in Action
22mBuilding Contextual Knowledge Graphs to Adapt, Evaluate, and Leverage LLMs
14mLLMs and Knowledge Graphs: A Case Study in Blue(y)
41mEnhancing LLMs with Graph Technology
20mPractical GraphRAG: Making LLMs smarter with Knowledge Graphs
29mPowering Legal Research with Knowledge Graphs and Large Language Models