Knowledge Base Grounding Score

The score is fundamentally dependent on correctly mapping textual mentions to unique, canonical identifiers within the knowledge graph. A claim about 'Paris' must be resolved to the Wikidata entity for the capital of France (Q90) and not Paris, Texas (Q126392). High grounding requires high-precision entity linking, as a mismatch at this stage invalidates the entire factual verification pipeline. The score penalizes ambiguous references that cannot be confidently resolved to a single node.

Beyond identifying entities, the score measures whether the stated relationship between them matches a defined property (predicate) in the knowledge graph. For the claim 'Tim Berners-Lee invented the World Wide Web', the system verifies if a directed edge labeled P61 (creator) or P800 (notable work) exists between the nodes Q80 and Q466 in Wikidata. The score reflects the semantic precision of the relationship, not just the co-occurrence of entities.

A robust grounding score incorporates temporal qualifiers to prevent factual drift. The claim 'Apple's CEO is Steve Jobs' would receive a low score because the P1308 (officeholder) edge for Q312 (Apple) includes start and end times, marking the statement as historically true but currently false. The score algorithmically applies temporal reasoning to ensure the claim is valid within the specified or implied timeframe, penalizing outdated facts.

The score applies a decay function based on the path length between entities in the knowledge graph. A direct, first-degree connection (e.g., 'The capital of France is Paris') receives a maximum score. A claim requiring multi-hop inference (e.g., 'The currency of the country that won the 1998 FIFA World Cup is...') receives a lower score due to the increased risk of compounding errors. The metric quantifies logical proximity, rewarding direct assertions over complex inferential chains.

Not all nodes and edges in a knowledge graph are created equal. The grounding score weights facts based on the provenance of the data. Statements sourced from a curated enterprise graph or a high-reliability Wikidata reference (e.g., a DOI link to a peer-reviewed paper) are weighted more heavily than those from crowd-sourced or unverified imports. The score integrates a meta-layer of source credibility, ensuring the knowledge base's own internal trust hierarchy is reflected in the final metric.

A high grounding score requires the absence of contradictory evidence within the knowledge base. The system actively searches for conflicting statements. If a claim asserts a fact, but the knowledge graph contains an explicit negation or a mutually exclusive property value for the same entity, the score is severely penalized. This transforms the metric from a simple lookup into an active verification process, ensuring logical consistency with the entire body of structured knowledge.

A composite metric that quantifies the degree of semantic and factual correspondence between a specific AI-generated statement and the content of its cited source. This score combines semantic similarity vectors with factual entailment ratios to produce a single alignment value.

• High alignment: The source directly supports the claim with matching terminology and data
• Low alignment: The source is topically related but does not substantiate the specific assertion
• Zero alignment: The citation is irrelevant or contradicts the claim

Alignment scoring is a critical input to the broader grounding score, as it validates that citations are not merely decorative but genuinely evidentiary.

The calculated probability that a cited source document logically supports or entails a specific claim made in AI-generated text, determined through natural language inference (NLI) models. This ratio moves beyond keyword matching to assess logical relationships.

• Entailment: The source text logically implies the claim is true
• Neutral: The source text neither confirms nor contradicts the claim
• Contradiction: The source text provides evidence against the claim

Modern NLI systems use transformer-based architectures fine-tuned on datasets like MultiNLI and FEVER to achieve high accuracy in entailment detection. This metric directly feeds into the Knowledge Base Grounding Score by quantifying evidentiary support.

A dynamic, interconnected model representing entities—authors, institutions, domains—and their trust relationships, used to propagate and calculate authority scores across a network. Similar to PageRank but applied to citation networks rather than hyperlinks.

• Nodes: Individual sources, authors, publications, and institutions
• Edges: Citation links, co-authorship, institutional affiliations
• Weights: Adjusted by factors like H-index, peer-review status, and retraction history

The graph enables transitive trust propagation: a source cited by multiple high-authority entities inherits credibility. This graph-based authority score is a foundational input to grounding calculations, ensuring that alignment with trusted nodes carries greater weight.

A measure of the completeness and logical validity of the path from an AI's output claim back through its citations to the foundational, verifiable data. This metric evaluates whether the citation chain is unbroken and each link is valid.

• Complete chain: Claim → Cited source → Source's references → Primary data
• Broken chain: Missing intermediate citations or reliance on unverifiable sources
• Circular chain: Sources citing each other without external grounding

Evidence chain integrity is essential for the grounding score because a claim aligned with a source that itself lacks proper provenance cannot be considered fully grounded. The Citation Chaining Protocol automates this recursive verification.

A verification technique that checks for agreement among multiple independent, high-quality sources to confirm a claim, increasing confidence through corroboration. This method reduces reliance on any single source and mitigates the risk of citing an outlier or erroneous publication.

• Strong consensus: Multiple Tier 1 sources independently confirm the claim
• Weak consensus: Agreement exists but among lower-tier or interdependent sources
• No consensus: The claim appears in only one source or sources conflict

Cross-reference consensus directly strengthens the grounding score by demonstrating that a claim is not merely aligned with one knowledge base entry but is reproducibly verifiable across the broader information ecosystem.

A temporal decay function applied to a citation's authority score, prioritizing recently published or updated sources to ensure information freshness. This weighting mechanism prevents grounding scores from being inflated by outdated but historically authoritative sources.

• Exponential decay: Authority diminishes rapidly for sources older than a domain-specific threshold
• Domain-specific windows: Medical literature may decay faster than historical scholarship
• Update refresh: A source's recency weight resets upon verified content revision

Recency weighting is critical for grounding scores in fast-moving fields where knowledge becomes obsolete quickly. A perfectly aligned but decade-old source may receive a lower grounding contribution than a recent, equally aligned publication.