Overview
Dong, Gabrilovich et al. (Google, KDD 2014) — Introduces Knowledge Vault (KV), a web-scale probabilistic knowledge base that fuses extractions from web content with prior knowledge from existing repositories. Substantially larger than any previously published structured knowledge repository.
Architecture: Three Components
| Component | Function |
|---|---|
| Extractors | Extract triples from web sources (text, tables, page structure, human annotations). Each assigns confidence scores. |
| Graph-based Priors | Learn prior probability of each triple based on existing KB (Freebase) |
| Knowledge Fusion | Compute probability of triple correctness based on agreement between extractors and priors |
Key Innovation
Analogous to speech recognition: combines noisy signals (web extractions) with prior models (existing KB). The prior model overcomes extraction errors AND source errors (e.g., web pages claiming incorrect facts).
Problem Statement
Existing KBs are far from complete — 71% of Freebase people lack place of birth, 75% lack nationality. Wikipedia growth has plateaued. Automated extraction from the full web is necessary but noisy. KV’s probabilistic approach manages this noise.
Technical Detail
Stores RDF triples with confidence scores. Fixed ontology (not Open IE). Language-independent: separates facts from lexical representation.
Namesake Note
This paper shares its name with our knowledge vault project but the concepts are relevant — probabilistic fusion of diverse information sources into a unified knowledge structure is exactly what research integration does.
Related
- neural-symbolic-machines — Later work combining neural and symbolic approaches for knowledge base QA
- ngram-machines-knowledge-qa — Related knowledge graph querying approach