Entity Linking

The foundational preprocessing step that identifies and classifies text spans into predefined categories such as person, organization, location, or date. Modern systems use transformer-based sequence labeling models like BERT or span-based architectures to detect entity boundaries before disambiguation begins.

• Token Classification: Assigns BIO (Begin, Inside, Outside) tags to each token
• Span Detection: Identifies contiguous token groups as candidate mentions
• Fine-Grained Typing: Extends beyond coarse types to labels like 'Politician' or 'PublicCompany'

Produces a shortlist of possible knowledge base entries for each detected mention. This retrieval step balances recall against computational efficiency, typically using alias tables, surface form dictionaries, and inverted indexes built from entity labels, redirects, and disambiguation pages.

• Alias Dictionary: Maps surface forms (e.g., 'USA') to candidate Q-Nodes (Q30)
• Prior Probability: Uses anchor text statistics from Wikipedia to rank candidates by commonness
• Name Variants: Handles acronyms, abbreviations, and multilingual labels

The core ranking step that selects the correct entity from the candidate set by measuring semantic coherence between the mention's surrounding text and the entity's knowledge graph description. State-of-the-art approaches use cross-encoder models that process the mention context and entity abstract jointly.

• Local Compatibility: Compares mention context words to entity description tokens
• Global Coherence: Ensures all linked entities in a document relate to a consistent topic
• Graph-Based Ranking: Uses PageRank-style algorithms over entity-entity relationship graphs

The critical capability to recognize when a mention refers to an entity that does not exist in the target knowledge base. Without nil prediction, systems force incorrect links, introducing noise. Modern linkers assign a NIL threshold score below which mentions remain unlinked.

• Out-of-KB Detection: Flags emerging entities not yet cataloged in Wikidata or DBpedia
• Confidence Thresholding: Rejects candidates when maximum similarity score falls below a calibrated cutoff
• Clustering: Groups unlinkable mentions that likely refer to the same new entity

Architectures that jointly optimize mention detection and entity disambiguation in a single model, eliminating error propagation between pipeline stages. Systems like GENRE use sequence-to-sequence models that generate entity names directly from input text, while autoregressive approaches output canonical URIs.

• Joint Optimization: Backpropagates through both detection and disambiguation losses
• Retrieval-Augmented: Combines dense passage retrieval with generative models
• Zero-Shot Linking: Links to entities never seen during training by leveraging entity descriptions

Standardized benchmarks measure linking quality using precision, recall, and F1 score computed at the mention level. The GERBIL platform provides a unified evaluation framework across multiple annotators and datasets like AIDA-CoNLL, MSNBC, and WebQuestions.

• Micro F1: Aggregates performance across all mentions equally
• Macro F1: Averages per-document performance to avoid bias toward long texts
• In-KB Accuracy: Measures correctness only for linkable mentions, excluding NIL cases

The core sub-task of entity linking that resolves ambiguous surface forms to a specific real-world entity. When a document mentions 'Paris', NED determines if it refers to the capital of France, the mythological figure, or a city in Texas by analyzing contextual clues like co-occurring entities and topic coherence. This is the critical decision point that separates simple named entity recognition from true semantic understanding.

A unique, persistent identifier (e.g., Q42 for Douglas Adams) assigned to every item in the Wikidata knowledge graph. Q-Nodes serve as the canonical URI for entity linking systems, providing a stable, language-independent anchor point. Unlike proprietary identifiers, Q-Nodes are human-curated and openly accessible, making them the preferred target for academic and enterprise knowledge graph population pipelines.

The NLP task of identifying all expressions in a text that refer to the same entity. Before linking 'Apple' to the company (Q312), the system must first resolve that 'it', 'the tech giant', and 'the Cupertino-based firm' all point to the same antecedent. Without robust coreference resolution, entity linking systems suffer from fragmented identity tracking and miss critical relationship signals.

An OWL property (owl:sameAs) used in RDF to explicitly declare that two different URIs represent the identical real-world entity. This is the fundamental mechanism for cross-source identity resolution, linking a DBpedia URI to its equivalent Wikidata Q-Node. Proper sameAs mapping is essential for knowledge graph consolidation and preventing entity fragmentation across linked data ecosystems.

A web service that programmatically matches local entity records against a remote knowledge base index. Services like the Wikidata Reconciliation API accept a name and optional metadata, then return ranked candidate matches with confidence scores. These APIs are the operational backbone of automated entity linking pipelines, enabling real-time disambiguation without maintaining a local copy of the entire knowledge graph.

A single, authoritative Uniform Resource Identifier designated to represent a specific entity. In entity linking, the canonical URI is the final output—the unambiguous identifier that consolidates all surface form variations. Using a canonical URI like http://www.wikidata.org/entity/Q312 for Apple Inc. prevents identity fragmentation and ensures all downstream systems reference the same entity consistently.