Entity Linking (EL)

The initial retrieval phase that narrows the search space from millions of Knowledge Base (KB) entries to a manageable list. This process combines surface form matching against a dictionary and alias tables to ensure high recall.

• Prior Probability (Commonness): Candidates are often ranked initially by the statistical likelihood P(e|m), derived from hyperlink counts in Wikipedia.
• Indexing: Systems like FAISS enable fast Approximate Nearest Neighbor (ANN) search over dense entity embeddings.
• Goal: Achieve >95% recall, ensuring the correct entity is in the candidate set for the next stage.

The core ranking stage where a model analyzes the semantic similarity between the mention's surrounding text and the candidate entity's description. This moves beyond static popularity to dynamic relevance.

• Bi-Encoder: Independently encodes the mention context and entity description into dense vectors for fast dot-product scoring.
• Cross-Encoder Reranker: A high-precision architecture that processes the concatenated mention and entity text jointly through full cross-attention, modeling fine-grained interactions.
• Entity Embeddings: Pre-computed dense vectors capturing the semantic properties of KB nodes.

The critical mechanism for identifying mentions that have no corresponding entry in the target Knowledge Graph. A system must not force a link to a wrong entity.

• Linking Confidence Score: A threshold is applied to the final probability score. If the score for the top candidate is below a certain value, the system predicts NIL.
• Out-of-KB Entity (OOKB): This handles real-world entities absent from the KB, preventing false positives.
• Methods: Some models train a dedicated binary classifier to distinguish linkable from unlinkable mentions.

A global optimization approach that resolves all mentions in a document jointly, rather than independently. It maximizes the semantic coherence among the final set of linked entities.

• Graph-Based Disambiguation: Constructs a graph of candidate entities and uses algorithms like Personalized PageRank to find the most central, coherent subgraph.
• Coherence Metrics: Measures the relatedness between candidate entities using Wikipedia Link-based Measure (WLM) or embedding similarity.
• Benefit: Resolves ambiguity where local context is insufficient, enforcing topical consistency.

Modern systems use transformer-based models that jointly perform mention detection and entity disambiguation, replacing traditional pipeline approaches.

• Entity-Aware BERT: Models like LUKE and KnowBERT natively integrate entity embeddings into the transformer's self-attention mechanism.
• Generative Entity Retrieval (GENRE): Frames EL as a sequence-to-sequence task, autoregressively generating the unique entity name token by token.
• BLINK: A two-stage model combining a fast Bi-Encoder for retrieval with a precise Cross-Encoder for final scoring, trained on billion-scale datasets.

A method for automatically creating large-scale, weakly labeled training data by heuristically aligning existing KB entries with text mentions, bypassing the need for expensive manual annotation.

• Anchor Text Alignment: Uses hyperlinks from Wikipedia as a primary source of positive examples, where the anchor text is the surface form and the link target is the entity.
• Noise Mitigation: Requires careful filtering to handle ambiguous anchors and pruning of over-generalized links.
• Scalability: Enables the creation of datasets with millions of examples, essential for training deep neural models.

The first stage of EL identifies the exact surface form—the literal text string—that refers to an entity. This involves segmenting spans like 'Apple' or 'Tim Cook' from unstructured text. The system must then generate a list of candidate entities for each mention, often using a dictionary built from anchor texts and redirect pages in Wikipedia. High recall at this stage is critical; a missed mention can never be linked.

Modern systems use a Bi-Encoder architecture for scalable candidate retrieval. The mention's context and each candidate entity's description are independently encoded into dense entity embeddings. A fast dot product scores thousands of candidates in milliseconds. This is often paired with a prior probability (or commonness) signal—the statistical likelihood that a surface form links to a specific entity based on Wikipedia hyperlink counts.

After fast retrieval, a Cross-Encoder Reranker performs high-precision scoring. Unlike a Bi-Encoder, it processes the concatenated mention text and single candidate entity jointly through full cross-attention. This captures deep semantic interactions but is computationally expensive, so it is applied only to the top-K candidates. This two-stage pipeline defines state-of-the-art models like BLINK.

Collective Entity Linking resolves all mentions in a document jointly rather than independently. The core assumption is that linked entities should be topically coherent. Graph-Based Disambiguation builds a subgraph of all candidate entities and uses algorithms like Personalized PageRank to find the most densely interconnected set. This global reasoning prevents local errors where one ambiguous mention is resolved incorrectly.

A robust system must perform NIL Prediction—correctly identifying that a mention has no corresponding entry in the target knowledge base. This prevents false links for Out-of-KB (OOKB) entities. The Linking Confidence Score, a value between 0 and 1, is thresholded. If the highest score falls below a tuned threshold, the system outputs NIL instead of forcing an incorrect link.

The standard benchmark is the AIDA CoNLL-YAGO dataset of Reuters news articles. The GERBIL Platform provides a unified evaluation framework to compare different EL tools fairly. Key metrics include precision, recall, and F1 score for both mention detection and linking accuracy. Systems like DBpedia Spotlight and generative models like GENRE are commonly evaluated on these benchmarks.