Entity Linking (EL)

The pipeline begins by identifying text spans that refer to entities. A Named Entity Recognition (NER) model classifies these spans into types (Person, Organization, Location).

• Boundary Detection: Accurately identifies the start and end of multi-word mentions like 'New York Stock Exchange'.
• Fine-grained Typing: Modern systems go beyond basic types to assign labels like 'Publicly Traded Company' to aid disambiguation.
• Nested Entity Handling: Advanced models resolve overlapping mentions, such as 'University of Cambridge' containing the location 'Cambridge'.

For each detected mention, the system retrieves a shortlist of possible matching entities from the target Knowledge Base (KB) (e.g., Wikidata).

• Surface Form Dictionary: Uses a pre-built index mapping common names (e.g., 'Apple') to candidate IDs (Q312, Q89).
• Dense Retrieval: Employs bi-encoder models to perform semantic search over entity descriptions, finding candidates even when exact names don't match.
• Alias Expansion: Incorporates redirects, acronyms, and common misspellings to maximize recall.

The core reasoning step where the system selects the correct entity from the candidate list by analyzing the surrounding text.

• Local Context Scoring: Compares the words around the mention to the entity's KB description using cross-encoder transformers.
• Entity Embedding Similarity: Measures the cosine similarity between the mention's contextual vector and pre-computed entity vectors.
• Prior Probability: Weighs the general popularity of an entity (e.g., 'Paris, France' is a stronger prior than 'Paris Hilton' in a geography article).

Instead of linking entities in isolation, this approach jointly disambiguates all mentions in a document to maximize topical agreement.

• Semantic Coherence: Ensures linked entities are related (e.g., 'Elon Musk' linked to 'Tesla' and 'SpaceX' is more coherent than linking to 'Apple').
• Graph-based Ranking: Constructs a graph of candidate entities and uses algorithms like PageRank to find the most densely connected subgraph.
• Nil Prediction: Critically identifies when a mention has no valid entry in the KB, preventing a false link to a similar but incorrect entity.

The final step assigns the resolved Canonical Entity Identifier (e.g., a Wikidata Q-ID) to the text mention, creating a machine-readable link.

• URI Assignment: Outputs a persistent, unique identifier that serves as the single source of truth.
• Confidence Scoring: Attaches a calibrated probability score (e.g., 0.98) to each link, allowing downstream systems to filter low-confidence predictions.
• Schema.org Alignment: Maps the linked entity to a Schema.org type for direct consumption by search engine crawlers and generative AI systems.

The foundational preprocessing step that locates and classifies textual mentions into pre-defined categories such as person, organization, or location. Entity Linking cannot proceed without accurate NER, as the system must first identify the boundaries of a mention before attempting to resolve it to a knowledge base entry. Modern approaches use transformer-based token classification with BIO tagging schemes.

The initial retrieval phase that produces a shortlist of possible knowledge base entries for a given mention. Techniques include:

• Surface form dictionaries built from anchor text and redirect pages
• Approximate nearest neighbor search over entity embedding spaces
• Alias tables that map common variations to canonical identifiers

Efficient candidate generation is critical for scaling to knowledge bases with millions of entities.

The core decision-making subtask that resolves which specific entity a mention refers to when multiple candidates share the same surface form. For example, distinguishing 'Paris' the city from 'Paris' the mythological figure. Modern NED systems compute a contextual similarity score between the mention's surrounding text and each candidate's knowledge base description, often using cross-encoder architectures.

The capability to correctly identify when a textual mention refers to an entity that does not yet exist in the target knowledge base. Without robust nil prediction, a linker will force a false positive match to an incorrect entity. This is implemented via a threshold on the linking confidence score or a dedicated unlinkable mention classifier trained on out-of-knowledge-base entities.

A global optimization approach that jointly disambiguates all entity mentions in a document simultaneously rather than independently. It maximizes semantic coherence by favoring sets of linked entities that have high topical agreement in the knowledge graph. This resolves ambiguous cases where a single mention is unclear in isolation but obvious when considering the document's overall subject.

A dense, low-dimensional vector representation of a knowledge graph entity that encodes its semantic properties and relational structure. Models like TransE, DistMult, and RotatE learn embeddings by optimizing for link prediction. These vectors enable fast similarity computation during candidate ranking and allow neural linkers to operate directly in a continuous semantic space.