Luna Dong, Principal Scientist, Amazon at MLconf Seattle 2017
summary
Knowledge Vault uses 15 extractors over 1B+ webpages to produce 3B+ triples with well‑calibrated probabilities (predicted 0.7 ≈ 70% accuracy). Knowledge‑Based Trust rates 119M webpages and 5.6M websites but covers ≤20% of sites because extraction favors known Freebase entities. A graphical model ties data items, sources and extractors to infer truth and source accuracy; single‑truth models reach ~0.992 precision/recall, multi‑type 0.91/0.98. Long‑tail work yields 2.2M triples for 100+ verticals and a product graph approach combines human‑in‑the‑loop labeling with hands‑off integration.
AI-written summary · about
chapters (6)
- 0:00Predicting triple correctness
- 1:58Knowledge-based trust
- 4:45Knowledge fusion model
- 7:33Applications and challenges
- 10:50Product graph development
- 14:35Questions and answers
related talks
25mEnhance LLM Applications' Trustworthiness and Explainability
30mBuilding a Conversational Knowledge Base
45mSafely Capture Institutional Knowledge Using
44mBuilding Knowledge Graphs
23mKnowledge Graphs and LLMs in Action
33mSmarter Industrial Fault Diagnosis with Knowledge Graphs and Large Language Models
46mBuilding Knowledge Graphs
38mBuilding a knowledge graph for climate policy