An Agentic Hybrid Top-Down and Bottom-Up Approach to Knowledge Graph Generation
The hosts discuss a paper from Malt on generating knowledge graphs from messy freelancer skill data. They explain the hybrid approach: anchoring recognized skills to Wikidata entities while using LLMs to discover and integrate new skills, with an agentic reflection loop for self-correction. Results show 77% resolution and 52% compression, with a focus on auditability and bias.