Entity Linking

The initial boundary detection phase that identifies text spans referring to entities. This step distinguishes between common words and proper nouns, handling complex cases like nested mentions and overlapping spans.

• Surface Form Extraction: Isolating the exact string used in text (e.g., 'Apple' in 'Apple announced a new chip').
• NIL Detection: Identifying mentions that have no corresponding entry in the target knowledge base, crucial for maintaining precision without forcing false matches.
• Overlap Resolution: Deciding boundaries for mentions like 'University of California, Berkeley' versus 'Berkeley'.

The recall-focused step that retrieves a set of possible knowledge base entries for a given mention. This phase balances efficiency against coverage, using alias tables and prior probability maps.

• Name Dictionary Lookup: Using pre-built mappings from surface forms to entity IDs, often sourced from Wikipedia redirects and disambiguation pages.
• Acronym Expansion: Resolving 'ML' to 'Machine Learning', 'Major League', or 'Mali' based on context.
• Search Engine Indexing: Leveraging inverted indexes over entity descriptions to find candidates via token overlap.

The core ranking step that scores each candidate based on its coherence with the surrounding text. Modern systems use dense retrieval and cross-encoders to model semantic compatibility.

• Local Context Scoring: Comparing the mention's surrounding words against each candidate's description embedding.
• Global Coherence: Maximizing the topical consistency across all linked entities in a document, ensuring 'Paris' links to the city if 'France' is also mentioned.
• Type Constraints: Filtering candidates by expected entity type (e.g., PER, ORG, LOC) derived from the mention's syntactic role.

The final resolution step that maps the selected candidate to its unique, stable identifier in the target graph. This creates the deterministic link required for downstream reasoning.

• Canonical ID Assignment: Linking to a persistent URI like wikidata:Q312 for 'Apple Inc.' versus wikidata:Q89 for the fruit.
• Provenance Tracking: Recording the confidence score and evidence used for the link to support auditability and debugging.
• Alias Absorption: Normalizing all surface form variations ('JFK', 'John F. Kennedy', 'President Kennedy') to a single node.

Modern neural approaches that jointly perform mention detection and entity disambiguation in a single pass, eliminating the error propagation common in traditional pipeline systems.

• Sequence-to-Sequence Generation: Autoregressively generating annotated text with entity boundaries and IDs directly from the raw input.
• Span-Based Classification: Enumerating all possible text spans and classifying each into an entity ID or a null class.
• Retrieval-Augmented Linking: Using a bi-encoder to embed the mention context and retrieve the top-k candidates from a dense entity index, then re-ranking with a cross-encoder.

Standardized measures to quantify linking system performance, focusing on the correctness of both the detected mention boundaries and the assigned knowledge base identifiers.

• Precision@K: The fraction of top-K ranked candidates that are correct, critical for systems where a downstream process re-ranks candidates.
• Linking F1: The harmonic mean of mention-level precision and recall, requiring both correct span detection and correct entity assignment.
• KB-Accuracy: The strict accuracy of the entity ID assigned, calculated only over correctly detected mentions to isolate disambiguation quality.

The foundational information extraction task that locates and classifies named entities in text into pre-defined categories. While NER identifies what a span of text is (e.g., 'Apple' is an ORG), Entity Linking goes further to determine which specific entity it refers to (e.g., Apple Inc. vs. Apple Corps).

• Input: Unstructured text
• Output: Labeled spans (Person, Org, Location)
• Relationship: NER is a prerequisite for Entity Linking

The NLP task of identifying all expressions that refer to the same real-world entity within a text. This includes linking pronouns ('she', 'it') and nominal phrases ('the company') back to their antecedents.

• Example: Resolving 'it' in 'Apple launched a new phone. It sold well.' to 'Apple'
• Impact: Without coreference resolution, entity linking systems miss implicit mentions and produce incomplete knowledge graph connections

The computational task of identifying which meaning of a polysemous word is intended in context. This is critical for entity linking when a mention is ambiguous.

• Example: Determining if 'Washington' refers to the state, the city, or the person
• Technique: Uses surrounding context words as features for classification
• Synergy: WSD resolves lexical ambiguity before entity linking maps the disambiguated sense to a knowledge base entry

The process of building structured semantic networks from unstructured data. Entity Linking is the bridge that connects textual mentions to nodes in these graphs.

• Nodes: Entities (people, places, concepts)
• Edges: Typed relationships (worksAt, locatedIn, foundedBy)
• Role of Entity Linking: Populates the graph by grounding textual references to canonical identifiers like Wikidata Q-IDs

The process of extracting specific attributes from a query to populate a predefined template. Entity Linking provides the canonical entity identifiers that slot filling uses to retrieve attributes from a knowledge base.

• Example Query: 'When was the CEO of Microsoft born?'
• Entity Linking: Resolves 'Microsoft' to Q2283 (Microsoft Corp)
• Slot Filling: Extracts the intent to find the birth date of the CEO attribute of that entity

The process of analyzing a query to determine its domain, temporal range, or constraints. Entity Linking anchors the query to specific real-world concepts, enabling precise scoping.

• Without Linking: 'Paris' could be about the city, the mythological figure, or Paris Hilton
• With Linking: Resolving to Q90 (Paris, France) scopes the search to geography, travel, or French politics
• Result: Dramatically reduced search space and improved precision