Neural Entity Linking

Unlike traditional systems that treat mention detection and entity disambiguation as separate pipeline stages, neural approaches use a shared transformer encoder to jointly represent both the textual mention context and the candidate entity descriptions. This allows the model to learn context-sensitive mention boundaries while simultaneously computing compatibility scores against knowledge base entries. The architecture typically employs a bi-encoder or cross-encoder design where mention representations and entity embeddings are projected into the same dense vector space for efficient nearest-neighbor retrieval.

• Topical coherence with surrounding entities
• Lexical overlap between context and entity description
• Type compatibility learned from training data
• Prior probability derived from entity popularity in the knowledge base

Modern neural EL systems employ a dual-encoder architecture where mention and entity encoders are trained jointly using a contrastive loss function. The model maximizes the similarity between a mention embedding and its correct entity while minimizing similarity with hard negative candidates. This enables dense passage retrieval at inference time, where candidate entities are fetched via approximate nearest neighbor search over pre-computed entity embeddings stored in a vector index, dramatically reducing the computational cost compared to exhaustive cross-encoding.

A critical advantage of neural entity linking is its ability to generalize to unseen entities not present in the training data. By encoding entity descriptions rather than relying on fixed lookup tables, the model can disambiguate mentions referring to newly added knowledge base entries without retraining. This zero-shot capability is essential for production systems that must handle evolving knowledge bases. The model achieves this by learning to map textual descriptions to the same semantic space as mention contexts, enabling open-domain linking to dynamic entity catalogs.

Advanced neural EL architectures incorporate graph neural networks to perform collective entity linking across all mentions in a document simultaneously. The model constructs a mention-entity graph where nodes represent mention-candidate pairs and edges encode pairwise compatibility based on knowledge base relationships and semantic similarity. Message passing through this graph allows the model to propagate disambiguation signals, ensuring that linked entities form a coherent topical cluster. This global approach resolves ambiguous cases where local context alone is insufficient.

Neural entity linking systems leverage multilingual transformer encoders to perform cross-lingual disambiguation, mapping mentions in one language to entities in a knowledge base primarily described in another. The shared multilingual representation space allows the model to link a mention like 'Angela Merkel' in German text to the same Wikidata entity as '安格拉·默克尔' in Chinese. This capability is powered by language-agnostic sentence embeddings trained on parallel corpora, enabling truly global entity resolution without per-language training pipelines.

The foundational upstream task that locates and classifies named entity mentions in text before linking can occur. Modern NER systems use transformer-based token classification (e.g., BERT with a CRF head) to identify spans and assign coarse types like PERSON, ORG, or GPE. Neural entity linking systems often jointly optimize NER and EL in a single end-to-end architecture, sharing contextualized representations to reduce cascading errors from pipeline approaches.

• Span detection: Identifying the exact character offsets of entity mentions
• Type classification: Assigning fine-grained types (e.g., "American jazz pianist" vs. "person")
• Joint modeling: Training NER and EL simultaneously to propagate gradients through both tasks

The initial retrieval phase that produces a shortlist of possible knowledge base entries for a given mention. Neural approaches replace traditional surface form dictionaries with dense bi-encoder architectures that embed mentions and entity descriptions into a shared vector space for approximate nearest neighbor search. This phase critically balances recall and efficiency—the candidate set must include the correct entity while remaining small enough for computationally expensive cross-encoder reranking.

• Alias tables: Precomputed mappings from surface forms to candidate KB IDs
• Dense retrieval: Using FAISS or ScaNN indexes over entity embeddings
• Recall@64: Standard evaluation metric measuring if the gold entity appears in the top-64 candidates

The core decision step that resolves which specific entity a mention refers to when multiple candidates share the same surface form. Neural NED uses cross-encoder transformers that jointly encode the mention-in-context and each candidate entity's description, producing a compatibility score. This replaces traditional feature engineering (TF-IDF overlap, prior probability, type compatibility) with learned semantic matching that captures subtle contextual cues.

• Polysemy resolution: Distinguishing "Apple" (company) from "apple" (fruit)
• Contextual compatibility: Measuring semantic coherence between mention context and entity description
• Nil prediction: Identifying mentions with no valid KB entry to prevent false positive links

A global optimization approach that jointly disambiguates all entity mentions in a document by maximizing the semantic coherence among the linked entities. Rather than making independent local decisions, collective methods construct a mention-entity graph and apply loopy belief propagation or graph neural networks to enforce topical consistency. This resolves ambiguous mentions by leveraging the disambiguated context of co-occurring entities.

• Coherence models: Measuring relatedness between candidate entities using KB links or embedding similarity
• Graph-based inference: Treating the document as a fully connected factor graph
• Joint optimization: Maximizing the product of local mention-entity scores and pairwise entity-entity coherence scores

Dense vector representations of knowledge graph entities that encode their semantic properties and relational structure. Models like TransE, RotatE, and ComplEx learn embeddings by optimizing scoring functions over knowledge graph triples (head, relation, tail). In neural entity linking, pre-trained entity embeddings provide a type-aware prior that helps the model understand that "Barack Obama" is semantically closer to "president" than to "actor" even before reading the mention context.

• Translational models: TransE represents relations as vector translations: h + r ≈ t
• Bilinear models: ComplEx uses complex-valued embeddings to capture asymmetric relations
• Pretraining: Embeddings learned on Wikidata or Freebase provide transferable entity knowledge

The task of identifying mentions of the same entity across separate documents, enabling knowledge base construction from heterogeneous sources. While entity linking maps mentions to a fixed KB, cross-document coreference clusters mentions without a predefined ontology, discovering new entities and aliases. Neural approaches use agglomerative clustering over mention embeddings with learned pairwise similarity functions.

• Mention clustering: Grouping all references to the same real-world entity across a corpus
• Canonicalization: Selecting the most representative name for each discovered cluster
• KB augmentation: Adding newly discovered entities and aliases to existing knowledge bases