Entity Linking

The core function of resolving a textual mention to a single, unambiguous entry in a Knowledge Base (KB). This requires analyzing the context of the mention against the properties and relationships of candidate entities.

• Surface Form: The literal text string (e.g., 'Washington').
• Candidate Generation: Retrieving a set of possible entities from the KB (e.g., George Washington, Washington D.C., University of Washington).
• Contextual Ranking: Using algorithms like TF-IDF or neural cross-encoders to score candidates based on surrounding text.

A two-stage pipeline often confused with a single task. Mention Detection (or Named Entity Recognition) identifies the text span, while Entity Linking assigns the identity.

• End-to-End Systems: Modern transformer models often perform both tasks jointly to reduce error propagation.
• NIL Prediction: The crucial ability to recognize when a mention has no corresponding entity in the target KB, preventing forced, incorrect mappings.

A global optimization approach that disambiguates all mentions in a document simultaneously, rather than independently. The decision for one entity influences the probability of others.

• Graph-Based Algorithms: Construct a graph of mentions and candidates, then run algorithms like PageRank to find the most coherent set.
• Semantic Relatedness: Measures like PMI (Pointwise Mutual Information) or pre-computed entity embeddings are used to quantify the compatibility between co-occurring entities.

The target database for linking is a critical architectural choice. It defines the universe of possible identities.

• Wikidata: The most common open-source KB, identified by Q-IDs (e.g., Q30 for the United States).
• Enterprise Knowledge Graphs: Proprietary graphs containing internal product codes, client IDs, and domain-specific ontologies.
• Alias Tables: Pre-built dictionaries mapping surface forms, acronyms, and common misspellings directly to canonical IDs for high-speed lookup.

Performance is measured by the system's ability to correctly resolve mentions to their true KB identifiers.

• Precision@K: The fraction of mentions for which the correct entity is in the top K candidates.
• Micro/Macro F1: Standard precision and recall calculated across all mentions, with micro-averaging weighting by mention frequency.
• In-KB vs. Out-of-KB Accuracy: Separate metrics to evaluate performance on known entities versus the system's ability to correctly abstain (NIL prediction).

Modern systems leverage transformer-based bi-encoders and cross-encoders to create dense vector representations of mentions and entities.

• Bi-Encoder: Encodes the mention context and entity description independently, allowing for fast, pre-indexed candidate retrieval via approximate nearest neighbor (ANN) search.
• Cross-Encoder: Concatenates the mention and entity text as a single input to a transformer, providing a richer, higher-accuracy relevance score at the cost of computational speed.