Entity Linking

The initial stage of the pipeline identifies text spans that refer to entities. This involves Named Entity Recognition (NER) to classify spans into types (Person, Organization, Location) and mention expansion to resolve coreferent pronouns (e.g., linking 'she' to 'Ada Lovelace'). Modern systems often use transformer-based token classifiers to jointly detect boundaries and types, handling nested and discontinuous mentions that traditional sequence labeling models miss.

This phase retrieves a set of plausible knowledge graph identifiers for each detected mention. Techniques include:

• Alias dictionaries: Mapping surface forms (e.g., 'MJ') to known referents using pre-built indexes from Wikipedia anchor texts.
• Acronym expansion: Resolving 'NLP' to 'Natural Language Processing' using domain-specific inventories.
• Fuzzy string matching: Generating candidates via Levenshtein distance or phonetic hashing to handle typographical errors. The goal is high recall—ensuring the true entity is in the candidate set—while pruning the search space for the disambiguation step.

The core ranking step selects the correct entity from the candidate set by measuring semantic coherence. A local compatibility score computes the cosine similarity between the mention's surrounding text embedding and the candidate entity's document embedding (e.g., its Wikipedia page vector). A global coherence model then maximizes the collective relatedness among all linked entities in a document, often using Graph Neural Networks or Loopy Belief Propagation to enforce the constraint that all mentioned entities should belong to a coherent semantic domain.

A critical and challenging sub-task is identifying mentions that have no corresponding entity in the target knowledge graph. A 'NIL' or unlinkable prediction prevents the system from forcibly mapping a novel or obscure entity to an incorrect identifier. This is typically implemented as an additional class in the ranking model, where the mention's contextual embedding is compared against a learned threshold or an outlier detection boundary in the entity embedding space. High-accuracy NIL detection is essential for maintaining knowledge base purity.

Legacy systems pipeline separate NER, candidate generation, and ranking models. Modern approaches use joint neural models that perform mention detection and entity disambiguation simultaneously. These architectures leverage cross-attention mechanisms between token representations and entity embeddings, allowing the model to directly compute a probability distribution over all entities in the knowledge graph for each token span. This eliminates cascading errors from upstream components and enables gradient flow through the entire linking process.

Entity Linking performance is quantified using standard information retrieval metrics computed at the mention level:

• Micro/Macro Precision & Recall: Measuring the fraction of correctly linked mentions.
• F1 Score: The harmonic mean of precision and recall, the primary benchmark metric.
• In-KB Accuracy: Evaluated only on mentions with a gold entity in the knowledge base, excluding NIL mentions.
• AIDA CoNLL-YAGO and GERBIL are the standard benchmarking datasets and platforms for reproducible evaluation.

The computational sibling of entity linking that focuses on deduplication and record merging. While entity linking connects a mention to a canonical knowledge base entry, entity resolution determines whether two records refer to the same real-world entity. Techniques include:

• Blocking: Grouping candidate pairs to reduce quadratic comparison costs
• Fuzzy string matching: Using Levenshtein distance or Jaccard similarity
• Probabilistic matching: Applying Fellegi-Sunter models for statistical linkage Modern systems often combine both tasks in a single pipeline.

A closely related subtask that specifically resolves ambiguous entity mentions to their correct referent. For example, disambiguating 'Washington' to the city, the state, or the person. Core approaches include:

• Local models: Using context windows around the mention
• Global models: Jointly disambiguating all mentions in a document for coherence
• Neural ranking: Fine-tuning BERT-based cross-encoders to score mention-entity pairs NED is often used interchangeably with entity linking in academic literature.

The structured target database that entity linking resolves mentions against. A knowledge graph stores entities as nodes and their relationships as edges, providing the unique identifiers (URIs) that linking systems output. Key characteristics:

• Ontology-driven: Entities are typed (Person, Organization, Location)
• Relationship-rich: Edges capture semantic connections like 'foundedBy' or 'locatedIn'
• Deterministic: Unlike vector stores, facts are explicit and queryable via SPARQL or Cypher Popular knowledge graphs include Wikidata, DBpedia, and enterprise-specific implementations.

The first-stage retrieval step in entity linking that produces a shortlist of possible entities for a given mention. This stage must balance recall against efficiency. Common techniques:

• Name dictionary lookup: Pre-built mapping from surface forms to entity IDs using anchor text from Wikipedia
• Alias tables: Indexing all known names, acronyms, and common misspellings
• Approximate nearest neighbor search: Using entity embeddings for fuzzy matching
• Prefix trees (tries): For real-time autocomplete-style candidate retrieval A typical candidate generator returns 5-20 candidates per mention for downstream ranking.

The second-stage scoring phase that selects the correct entity from the candidate set. Modern systems use transformer-based architectures:

• Cross-encoders: Concatenate mention context with entity description and pass through BERT for a relevance score
• Bi-encoders: Encode mentions and entities separately, then compute cosine similarity for efficiency
• Entity embeddings: Pre-computed vector representations from graph neural networks or knowledge graph embeddings
• Type constraints: Filtering candidates by expected entity type (PER, ORG, LOC) to improve precision The ranking model is the core differentiator in production entity linking systems.