Entity Reconciliation

The core mechanism maps a local entity record to a globally unique, persistent identifier (URI) in an authoritative knowledge base like Wikidata or the Google Knowledge Graph. This process uses the sameAs property in Schema.org to explicitly assert equivalence, consolidating all authority signals to a single, non-negotiable node. Without a canonical ID, an entity remains ambiguous and its authority is fragmented across multiple interpretations.

Reconciliation engines rarely rely on exact string matching. They employ fuzzy string algorithms (Levenshtein distance, Soundex) and semantic vector similarity to overcome variations:

• Typographical errors: 'Mcdonalds' vs. 'McDonald's'
• Aliases: 'The Big Apple' vs. 'New York City'
• Multilingual variations: 'Munich' vs. 'München' This probabilistic layer scores candidate matches, returning a confidence threshold rather than a binary result.

A single name often refers to multiple distinct entities. Reconciliation resolves this by analyzing contextual attributes surrounding the mention:

• Entity type constraints: Is 'Mercury' a planet, an element, or a car brand?
• Co-occurring entities: A mention of 'Paris' alongside 'Eiffel Tower' and 'Seine' strongly signals the French capital.
• Topical domain: In a medical text, 'ALS' likely means Amyotrophic Lateral Sclerosis, not Advanced Life Support. This step is critical for high-precision knowledge graph grounding.

To break ties between high-scoring candidates, reconciliation services cross-reference disambiguating properties from the knowledge base. For a person entity, this might include:

• birthDate and birthPlace
• occupation or employer
• external identifiers (VIAF, ISNI, ORCID) A mismatch in a single high-weight property (e.g., a birth date off by a century) can correctly disqualify an otherwise high-similarity candidate.

Reconciliation operates in two primary modes:

• Batch Processing: Uploading entire datasets (CSV, JSON) to a service like OpenRefine or a custom reconciliation API to process millions of records asynchronously.
• Streaming/Real-time: Using a low-latency API endpoint to reconcile entities on-the-fly during content creation or user input, ensuring data is canonicalized at the point of entry. Both modes are essential for maintaining a continuously clean, linked data architecture.

Every candidate match is assigned a numerical confidence score (typically 0.0 to 1.0). System architects must define strict operational thresholds:

• Automatic acceptance: Score > 0.95 triggers an automated link.
• Human review queue: Score between 0.7 and 0.95 is flagged for manual curation.
• Rejection: Score < 0.7 creates a new, unlinked entity or is left unresolved. This triage system balances automation efficiency with the absolute precision required for authoritative knowledge graphs.