Reading The Web with Learned Syntactic-Semantic Inference Rules

Ni Lao (CMU), Amarnag Subramanya (Google), Fernando Pereira (Google), William W. Cohen (CMU).

Key Contribution

Web-scale implementation of PRA that combines knowledge base relations with syntactic patterns from parsed text for relation extraction. Scales to Freebase + 60 million parsed documents.

Approach

• Distributed implementation of path-constrained random walk model
• Learns inference rules that combine KB edges with syntactic-semantic patterns from text
• Rules are defeasible (e.g., “born in city in country → national of country” may fail for naturalization)
• Multiple imperfect rules are combined to yield useful KB completions

Scale

• Knowledge base: Freebase
• Text corpus: 60 million parsed web documents
• Distributed computation for tractable inference at this scale

Significance

Bridges structured KB reasoning (PRA) with unstructured text extraction — a key step toward knowledge base completion from heterogeneous sources.

Connection to Vault

Direct extension of the core Path Ranking Algorithm (Lao & Cohen, 2010-2011). Precursor to the Knowledge Vault project which further scaled KB completion.