Entity Linking

Entity linking is the natural language processing task of mapping mentions of entities in unstructured text to their corresponding unique identifiers in a target knowledge base such as Wikidata or a domain ontology. It combines candidate generation, which retrieves plausible referents for a surface form, with entity disambiguation, which selects the correct referent using contextual, semantic, and popularity signals. The task resolves ambiguity where one surface form may denote many entities and one entity may be expressed by many surface forms. Entity linking is a foundational step in transforming free text into structured, machine-readable knowledge.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:hasPart ai:MentionDetection))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:hasPart ai:CandidateGeneration))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:hasPart ai:EntityDisambiguation))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:hasPart ai:NILDetection))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:hasPart ai:AliasExpansion))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:hasPart ai:GlobalCoherenceModel))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:hasPart ai:CandidateRanker))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:hasPart ai:EntityIndex))

Dependency Relationships

SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:requires ai:NamedEntityRecognition))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:requires ai:KnowledgeGraph))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:requires ai:KnowledgeBase))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:requires ai:AliasDictionary))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:depends_on ai:Embedding))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:depends_on ai:TransformerArchitecture))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:depends_on ai:CoreferenceResolution))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:depends_on ai:Tokenisation))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:depends_on ai:ApproximateNearestNeighbourSearch))

Capability Relationships

SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:enables ai:EntityResolution))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:enables ai:KnowledgeGraphConstruction))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:enables ai:QuestionAnswering))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:enables ai:SemanticSearch))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:enables ai:RelationExtraction))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:enables ai:KnowledgeBasePopulation))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:enables ai:EventExtraction))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:supports ai:TextMining))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:supports ai:InformationRetrieval))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:supports ai:DialogueSystems))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:supports ai:ContentRecommendation))

Implementation Relationships

SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:implements ai:NamedEntityDisambiguation))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:implements ai:Wikification))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:uses ai:BiEncoder))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:uses ai:CrossEncoder))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:uses ai:TrieDecoding))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:uses ai:DenseRetrieval))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:uses ai:Wikidata))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:uses ai:BERT))

Reduction Relationships

SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:reducesTo ai:InformationExtraction))
SubClassOf(ai:EntityLinking
  ObjectSomeValuesFrom(ai:reducesTo ai:EntityDisambiguation))
SubClassOf(ai:CandidateGeneration
  ObjectSomeValuesFrom(ai:reducesTo ai:InformationRetrieval))
SubClassOf(ai:EntityDisambiguation
  ObjectSomeValuesFrom(ai:reducesTo ai:Ranking))
SubClassOf(ai:GlobalCoherenceModel
  ObjectSomeValuesFrom(ai:reducesTo ai:GraphInference))
SubClassOf(ai:NILDetection
  ObjectSomeValuesFrom(ai:reducesTo ai:BinaryClassification))

About

Entity linking (also called named entity disambiguation, NED, or “wikification” when the target is Wikipedia) is the problem of grounding free-text entity mentions to canonical entries in a structured Knowledge Base. The discipline emerged from early research on word-sense disambiguation in the 1990s and was crystalised as a distinct NLP task by the TAC Knowledge Base Population (KBP) shared tasks from 2009 onwards. Unlike Named Entity Recognition, which merely labels a span as PERSON, ORGANISATION, LOCATION, or similar, entity linking must select the specific unique identifier — such as Q7251 in Wikidata for Alan Turing — from potentially millions of candidates. The challenge is compounded by surface form variation: an entity may appear under its canonical name, an abbreviation, a pronoun, or a colloquial alias, while the same surface form may be genuinely ambiguous between many distinct entities in the Knowledge Graph. Early entity linking systems in the 2000s used Wikipedia anchor statistics: anchor text hyperlinks in Wikipedia articles provide a natural mapping from surface forms to the articles they link to, and the prior probability P(entity | surface form), estimated from anchor counts, proved to be a strong baseline. Milne and Witten (2008) formulated entity linking as a machine learning problem combining surface-form priors, contextual similarity (based on the bag-of-words overlap between context and Wikipedia article text), and graph-based coherence signals (relatedness between candidate entities estimated from their Wikipedia link co-occurrence). Their approach, published in CIKM 2008, remained competitive for several years and demonstrated the value of combining local and global signals.

The introduction of distributed word representations dramatically changed the landscape. Yamada et al. (2016) trained entity embeddings jointly with word embeddings such that the embeddings of entity names were close to embeddings of contexts in which they appear, enabling purely distributional disambiguation without explicit feature engineering. Deep learning architectures — convolutional neural networks initially, then bidirectional LSTMs and attention mechanisms — improved mention detection and candidate ranking further. The real paradigm shift, however, came with the Transformer Architecture and the rise of pre-trained language models. Wu et al. (2020) introduced BLINK, a Bi-Encoder architecture built on BERT that encodes mentions and entity descriptions independently into a shared dense vector space. Candidate retrieval is performed via approximate nearest-neighbour search over a dense index (using FAISS), and a cross-encoder then re-ranks the top-64 candidates. BLINK demonstrated that scalable entity linking to millions of entities was achievable with BERT-level representations, and established the retrieve-then-rerank paradigm as the dominant architecture for several subsequent years.

De Cao et al. (2021) fundamentally reframed the problem with GENRE (Generalised Autoregressive Entity Retrieval), which treats entity linking as a sequence-to-sequence generation task. Instead of indexing all entities and performing retrieval, GENRE autoregressively generates the entity name character-by-character using a constrained beam search over a prefix trie, guaranteeing that only valid entity names in the target base can be output. This eliminates the need for any pre-built index and enables zero-shot transfer to new domains or languages simply by changing the entity name set. The multilingual extension mGENRE (De Cao et al., 2022) applies the same approach across 100+ languages, demonstrating remarkable cross-lingual transfer. As of 2024–2025, the dominant approaches are either generative (GENRE-style), retrieve-and-rerank (BLINK-style), or hybrid frameworks that combine retrieval candidates with generative reranking. The OneNet framework (EMNLP 2024) demonstrated that combining contextual priors and knowledge base statistics through few-shot LLM prompting can outperform prior state-of-the-art across seven benchmark datasets without task-specific fine-tuning, pointing toward a future where entity linking is an emergent capability of large language models rather than a specialist pipeline stage.

Components / Architecture

A complete entity linking system consists of the following components, which may be implemented as separate modules in a pipeline or jointly in an end-to-end model:

• Mention Detection: Identifies entity-bearing spans in text. Can be a dedicated Named Entity Recognition module producing (start, end, type) tuples, a span proposal network that scores all possible spans, or an integrated model that jointly detects and links. Mention detection quality directly bounds recall: missed mentions cannot be linked. Recent end-to-end models (ELQ, ReFinED) perform mention detection and disambiguation in a single forward pass, with cross-attention between all spans and all entity descriptions.

• Alias Dictionary and Entity Index: A pre-built lookup table mapping surface forms and aliases to candidate entity identifiers. For Wikipedia-based systems, article titles, redirect page titles, and anchor text provide the alias vocabulary. For Wikidata, aliases, labels, and description strings in all supported languages are crawled and indexed. The alias dictionary is typically stored as an inverted index mapping surface form → {(entity_id, prior_prob), …} sorted by prior probability.

• Dense Retrieval / Bi-Encoder: A dual-tower Transformer Architecture (based on BERT or similar encoders) that maps mentions-in-context and entity descriptions to the same dense vector space, enabling fast approximate nearest-neighbour retrieval with FAISS or ScaNN. The mention encoder takes the mention span and its surrounding context tokens as input; the entity encoder takes the entity title and description text. During inference, the entity side is pre-encoded and indexed; only the mention side is computed dynamically. This is the dominant candidate generation mechanism in systems such as BLINK, DPR-EL, and ZESHEL.

• Cross-Encoder Re-ranker: A single Transformer Architecture model that concatenates the mention context with each candidate entity description and produces a compatibility score. Cross-encoders achieve higher accuracy than bi-encoders because they model interactions between mention and entity tokens via full self-attention, but they cannot scale to full-base retrieval: typically applied only to the top-64 or top-100 bi-encoder candidates. Cross-encoders can be implemented as BERT-based binary classifiers (mention-entity pair is / is not the correct link) or as sequence-to-sequence models scoring entity names.

• Generative Decoder with Trie Decoding: A seq2seq model (BART, T5, or GPT-style decoder) that autoregressively generates the canonical entity name or identifier. Constrained decoding via a prefix trie guarantees that intermediate generation prefixes are always consistent with at least one valid entity name, preventing hallucination of non-existent entities. GENRE uses BART fine-tuned with constrained beam search over a Wikidata/Wikipedia title trie. Generative models naturally handle zero-shot settings where no candidate dictionary exists.

• Global Coherence Module: An optional component that scores the joint assignment of all mention-entity pairs in a document, penalising mutually inconsistent combinations. Implemented historically as Loopy Belief Propagation over a Markov Random Field (Milne & Witten, Ratinov et al.), or as iterative graph neural network message passing (ELMO), or as cross-mention attention (recent transformer-based collective EL approaches). The coherence signal captures, for example, that “Apple” near “iPhone” and “Tim Cook” should all link to the company rather than the fruit or a different organisation.

• NIL Classifier: A threshold or learned classifier that determines when a mention has no referent in the target base (NIL / out-of-KB decision). Crucial for open-world entity linking — in biomedical and emerging-entity settings, a significant fraction of mentions refer to entities not in any existing knowledge base. NIL decisions require either learning a score threshold or training an explicit NIL classifier that distinguishes between in-KB and out-of-KB mentions based on the distribution of candidate scores.

  Formal Algorithm

  Given a document D containing a sequence of tokens, and a knowledge base KB of entities E = {e₁, e₂, …, eₙ} each with canonical identifiers and description strings:

  1. Mention detection: Extract set M = {m₁, m₂, …, mₖ} of candidate entity mention spans from D, where each mᵢ = (start_i, end_i, surface_form_i, type_i).

  2. Candidate generation: For each mention mᵢ, retrieve candidates Cᵢ ⊆ E:

  • From alias dictionary: Cᵢ_alias = AliasLookup(surface_form_i) → top-30 by prior P(eⱼ | surface_form_i)

  • From dense retrieval: Cᵢ_dense = ANN(encode_mention(mᵢ, context(mᵢ, D)), entity_index, K=64)

  • Combined: Cᵢ = Cᵢ_alias ∪ Cᵢ_dense (union, deduplicated) 3. Local scoring: For each (mᵢ, eⱼ ∈ Cᵢ), compute compatibility score:

  • sᵢⱼ = CrossEncoder(context(mᵢ, D), description(eⱼ)) ∈ ℝ 4. Global disambiguation (optional): Solve:

  • argmax_{e₁ ∈ C₁, …, eₖ ∈ Cₖ} Σᵢ sᵢ(eᵢ) + Σᵢ<ⱼ coherence(eᵢ, eⱼ)

  • where coherence(eᵢ, eⱼ) = entity-entity relatedness from KB link co-occurrence

  • Solved by beam search, loopy BP, or greedy approximation 5. NIL decision: For any mᵢ where max_{eⱼ ∈ Cᵢ} sᵢⱼ < τ (threshold), assign NIL. 6. Output: Return mapping {mᵢ → eᵢ*} ∪ {mᵢ → NIL} for all detected mentions.

    Major Families and Variants

• Retrieval-based (Bi-Encoder + Cross-Encoder): BLINK (Wu et al., 2020), DPR-EL, ZESHEL. Fast retrieval-then-rerank. Current state-of-the-art on CoNLL-AIDA and TAC-KBP. Weakness: requires a pre-built entity index; weaker at long-tail entities with sparse description text.

• Generative (Autoregressive): GENRE (De Cao et al., 2021), mGENRE (De Cao et al., 2022), EntGPT (2024). No candidate index needed; naturally zero-shot; slower inference than retrieval but stronger on cross-lingual and zero-shot benchmarks. Trie decoding prevents invalid outputs.

• Retrieve-and-Generate (Hybrid): OneNet (EMNLP 2024), combining retrieval-based prior knowledge with generative LLM prompting; achieves best results across seven benchmarks without any task-specific fine-tuning. SynCABEL (2025) is a hybrid that uses synthetic data augmentation for biomedical EL.

• End-to-End: ELQ (Li et al., 2020), GENRE-end2end, ReFinED (Ayoola et al., 2022) — jointly model mention detection and disambiguation in a single forward pass. ReFinED additionally introduces fine-grained entity typing as an auxiliary task to improve disambiguation.

• Collective / Global: AIDA (Hoffart et al., 2011), ELMO (global EL via graph neural message-passing). These approaches jointly resolve all mentions in a document using coherence constraints, bridging to Coreference Resolution.

• Zero-Shot / Domain-Adaptive: Logeswaran et al. (2019) first demonstrated zero-shot entity linking. BOOTLEG (Orr et al., 2021) specifically targets the long tail. SynCABEL (2025) addresses biomedical EL with synthetic contextualised augmentation for target domains without labelled data.

• Multilingual: mGENRE, XEL — entity linking across 100+ languages from a single model. Crucial for European and multilingual knowledge graph construction.

• Biomedical-Specific: MedLinker, BioSyn, Hybrid X-Linker (2024), SynCABEL (2025). These specialise in linking to UMLS, MeSH, SNOMED CT, and Gene Ontology, handling the distinctive properties of biomedical text (abbreviation overloading, species variation, novel entities not yet in any ontology).

  Use Cases and Major Application Domains

• Knowledge Graph Construction and Population: Automated ingestion of news, scientific papers, and enterprise documents into structured knowledge graph stores. Google’s Knowledge Vault pipeline, Freebase (now absorbed into Wikidata), and DBpedia Spotlight all rely on entity linking to populate and maintain large-scale knowledge bases. Wikidata editors increasingly use entity linking tools to identify and cross-reference new contributions automatically. The OpenKE framework (Zhu et al., 2018) provides embedding-based tooling for knowledge graph completion that relies on linked entities as its atomic unit.

• Question Answering and Retrieval-Augmented Generation: Modern open-domain Question Answering systems — including retrieval-augmented generation pipelines that power commercial conversational AI assistants — depend on entity linking to anchor retrieved passages to known entities, enabling faithful grounding of generated answers. Without EL, an LLM may retrieve a passage about “Apple (company)” when asked about “Apple (fruit)” or vice versa. Entity linking also enables precise slot-filling: given “Who founded Apple?”, the system must link “Apple” to Q312 (Apple Inc.) before retrieving the answer.

• Semantic Search and Recommendation: Search engines including Google, Bing, and DuckDuckGo use entity linking to transform keyword queries into entity-aware structured queries, enabling Knowledge Panel and Featured Snippet features. Recommendation systems in streaming platforms use entity linking over user reviews and watch histories to build entity-level preference profiles that improve recommendation diversity and serendipity beyond item-level collaborative filtering.

• Biomedical and Clinical NLP: Linking clinical notes to standardised ontologies — UMLS, MeSH, SNOMED CT, Gene Ontology — is critical for cohort discovery, pharmacovigilance, adverse event surveillance, and drug target identification. EL systems trained on PubMed and MIMIC corpora are deployed in hospital informatics pipelines across the NHS, NHS Digital, and Health Data Research UK (HDRUK) federated data platforms. The NHS Federated Data Platform, piloted from 2023 and scaling through 2025–2026, relies on entity linking to map clinical narrative text to SNOMED CT identifiers for population-level analytics.

• Legal and Financial Text Mining: Contracts, regulatory filings, court decisions, and earnings reports contain numerous entity mentions (companies, legislation, case references, financial instruments) that must be grounded to canonical identifiers for compliance monitoring, regulatory reporting, and due diligence workflows. HMRC, the FCA, and the Bank of England have invested in NLP pipelines that include entity linking to cross-reference financial entity mentions in regulatory submissions to LEIs (Legal Entity Identifiers) and GLEIF registry entries.

• Cross-Lingual Data Integration and Intelligence: Multilingual entity linking enables Record Linkage across databases in different languages, a critical capability for international intelligence sharing, humanitarian data coordination, and scientific dataset integration. The European Open Science Cloud (EOSC) uses entity linking to interlink research data repositories across member states speaking different languages.

• Content Tagging and Journalism: Automated entity linking over news articles, broadcast transcripts, and social media feeds enables real-time topic tracking, cross-publication story linking, and the generation of structured news summaries. The BBC, Reuters, and Associated Press have deployed entity linking at scale over their editorial archives to enable structured content discovery and licensing.

  Academic Context

  Entity linking as a formal shared task was established by the NIST Text Analysis Conference Knowledge Base Population (KBP) track, launched in 2009. The TAC-KBP entity linking task provided the first standardised benchmark: given a query entity name and a reference document, link it to a Wikipedia article or assign NIL. Early leading systems combined surface-form priors with contextual overlap features (Milne and Witten, 2008; Cucerzan, 2007; Ratinov et al., 2011). The shift to neural methods began with word embedding approaches (Sun et al., 2015; Yamada et al., 2016) and accelerated dramatically with transformer-based bi-encoders and cross-encoders (Gillick et al., 2019; Wu et al., 2020 — BLINK). The GENRE paper (De Cao et al., 2021) established autoregressive entity retrieval as an alternative paradigm, immediately adopted for zero-shot and multilingual settings. The ELQ system (Li et al., 2020) demonstrated end-to-end linking for open-domain question answering. The field then diversified: zero-shot EL (Logeswaran et al., 2019), entity linking for question answering (EntQA, Zhang et al., 2022), long-tail EL (BOOTLEG, Orr et al., 2021), multilingual EL (Upadhyay et al., 2018; mGENRE, De Cao et al., 2022), biomedical EL (BioSyn, Biosyn-BERT, SynCABEL), and LLM-based EL (EntGPT, 2024; OneNet, 2024).

  Key research groups include: University of Edinburgh NLP group (EdinburghNLP, HERD project — entity linking in historical and cultural heritage corpora), University College London NLP and clinical text mining groups (clinical EL to SNOMED CT via collaboration with UCLH), the University of Washington (Allen School — Luke Zettlemoyer’s group; ELQ, DPR-EL), Meta AI Research (BLINK, GENRE, mGENRE, EntGPT), IBM Research (TAC KBP systems), Allen Institute for AI (ELQ, BOOTLEG), Seoul National University (BioSyn), and the Alan Turing Institute (UK-specific NLP including EL over cultural heritage collections at the British Library and National Archives). Primary benchmark datasets: CoNLL-AIDA (Hoffart et al., 2011; Conll 2003 NER data with Wikipedia links), TAC-KBP 2009–2017, MSNBC/OKE, ZESHEL (zero-shot benchmark), MedMentions (biomedical), BC5CDR.

  Current Landscape (2026)

  As of mid-2026, entity linking has been substantially reshaped by the rise of large language models (Large Language Models). GPT-4, Claude 3, Llama 3, and Gemini can perform entity linking in zero-shot settings via structured prompting, delivering reasonable performance without any EL-specific training. While these approaches do not yet uniformly outperform fine-tuned specialist models on standard benchmarks, they offer remarkable flexibility: no alias dictionary, no entity index, and no training data are required for a new domain or language — only entity descriptions. The OneNet framework (EMNLP 2024) formalised this prompt-based approach, combining contextual cues and entity prior knowledge in a structured LLM prompt, and outperforming prior state-of-the-art across seven benchmark datasets without fine-tuning. SynCABEL (arXiv:2601.19667, 2025) addresses the persistent challenge of limited annotated biomedical EL data by generating synthetic contextualised training examples, achieving state-of-the-art performance on MedMentions with less than 5% of original annotated data. The Guiding LLMs for Biomedical Entity Linking paper (EMNLP Findings 2025) showed that targeted prompting strategies that leverage ontology structure outperform generic LLM prompting for clinical entity linking. Commercial entity linking APIs — Google Natural Language API, AWS Comprehend (DetectEntities), Microsoft Azure AI Language (recognise entities with knowledge base links) — are now standard components of enterprise data ingestion pipelines. In the NHS and UK biomedical sector, SNOMED CT entity linking is being automated through the NHS Federated Data Platform and the NHS England AI Lab. The integration of EL with retrieval-augmented generation (RAG) pipelines represents the most active commercial direction: linking extracted entities to KB descriptions and then using those descriptions to augment LLM context windows enables more factual, hallucination-resistant responses.

  UK Context

  Entity linking research in the UK has several well-established centres. The University of Edinburgh Institute for Language, Cognition and Computation (ILCC) and EdinburghNLP group have contributed to EL for cultural heritage and multilingual settings, with the HERD project applying EL to link person and place mentions in digitised historical newspaper archives to the Dictionary of National Biography and Getty Geographic Names Thesaurus. University College London’s NLP group has worked on clinical entity linking in collaboration with UCLH (University College London Hospitals), linking clinical text to SNOMED CT and ICD-10 codes within the UCLH-GOSH Digital Research Environment. The University of Manchester’s Bio-Health Informatics Group (BHI), in the NIHR Manchester Biomedical Research Centre, applies biomedical EL to cohort discovery and pharmacovigilance using Clinical Practice Research Datalink (CPRD) and Greater Manchester Care Record data. King’s College London’s Department of Informatics, in partnership with King’s College Hospital and the South London and Maudsley NHS Trust, has deployed entity linking pipelines for mental health text mining in the CRIS (Clinical Record Interactive Search) system.

  The Alan Turing Institute, based at the British Library in London, funds cross-institutional projects including Living with Machines (LwM), which applied NLP and entity linking at scale to digitised nineteenth-century British newspaper archives from the British Library’s collections, linking historical entities to Wikidata and the DNB. The University of Sheffield NLP group, home of the GATE (General Architecture for Text Engineering) framework which has long supported entity linking through the ANNIE annotation pipeline, remains an important centre for industrial-scale NLP deployments. In the North of England, the University of Leeds Medical Informatics group and the University of Newcastle’s Population Health Science Institute apply EL to primary care records (CPRD, QResearch) for large-scale epidemiological studies linking patient-reported symptoms and diagnoses to SNOMED CT and Read Codes. The NHS England AI Lab, headquartered in Leeds, funds EL tooling development as part of the NHS National Language Strategy.

  Future Directions (2026-2030)

• LLM-native EL with Implicit KB: Large language models trained on Wikipedia and Wikidata text implicitly learn entity representations; future fine-tuning approaches will enable EL as a direct structured generation capability without a separate entity index, maintained through continual learning updates as the knowledge base evolves.

• Streaming Knowledge Bases and Real-Time Linking: Current EL systems link to static snapshots of Wikidata or Wikipedia; production systems in news and financial analytics need EL against continuously updated knowledge graphs, requiring streaming retrieval index updates and online adaptation of disambiguation models.

• Multi-KB Disambiguation: Enterprise deployments require linking to internal proprietary knowledge bases alongside public ones, raising challenges around disambiguation when mentions could plausibly refer to entities in different bases, and privacy-preserving federated learning for cross-enterprise EL.

• Cross-Modal and Multimodal EL: Extending entity linking to images, audio, and video — grounding visual entity mentions (faces, logos, locations visible in images) to the same Knowledge Graph identifiers used for text — enabling truly multimodal knowledge graph construction. Early work on visual entity linking (linking faces to person entities, logos to company entities) is progressing through 2025–2026.

• Explainable and Auditable Linking: As EL feeds downstream decision-making in clinical, legal, and financial domains, regulatory requirements (EU AI Act, UK AI Liability Directive in prospect) will mandate that disambiguation decisions can be explained and audited. Future models will surface supporting evidence for each linking decision, including the specific context passages and entity description fragments that drove the choice.

• Quantum-Accelerated Retrieval: Quantum nearest-neighbour algorithms could dramatically accelerate the dense retrieval step for web-scale knowledge bases containing hundreds of millions of entities, a direction being explored at UK quantum computing centres including the National Quantum Computing Centre at Harwell and in IBM’s quantum network partnerships with UK universities.

  Research and Literature

  1. Milne, D. & Witten, I. H. (2008). “Learning to link with Wikipedia.” ACM CIKM 2008, 509–518. https://doi.org/10.1145/1458082.1458150
  2. Cucerzan, S. (2007). “Large-scale named entity disambiguation based on Wikipedia data.” EMNLP-CoNLL 2007, 708–716.
  3. Ratinov, L. et al. (2011). “Local and global algorithms for disambiguation to Wikipedia.” ACL 2011, 1375–1384.
  4. Shen, W., Wang, J. & Han, J. (2015). “Entity linking with a knowledge base: Issues, techniques, and solutions.” IEEE Transactions on Knowledge and Data Engineering, 27(2), 443–460.
  5. Yamada, I. et al. (2016). “Joint learning of the embedding of words and entities for named entity disambiguation.” CoNLL 2016, 250–259.
  6. Phan, M. C. et al. (2019). “Pair-linking for collective entity disambiguation: Two could be better than all.” IEEE Transactions on Knowledge and Data Engineering, 31(7), 1383–1396.
  7. Wu, L. et al. (2020). “Scalable zero-shot entity linking with dense entity retrieval (BLINK).” EMNLP 2020, 6397–6407. https://arxiv.org/abs/1911.03814
  8. De Cao, N. et al. (2021). “Autoregressive entity retrieval (GENRE).” ICLR 2021. https://arxiv.org/abs/2010.00904
  9. De Cao, N. et al. (2022). “Multilingual autoregressive entity linking (mGENRE).” Transactions of the ACL, 10, 274–290.
  10. Ayoola, T. et al. (2022). “ReFinED: An efficient zero-shot-capable approach to end-to-end entity linking.” NAACL 2022. https://arxiv.org/abs/2207.04108
  11. Li, Z. et al. (2020). “EFFICIENT ONE-PASS END-TO-END ENTITY LINKING FOR QUESTIONS (ELQ).” EMNLP 2020, 6433–6444.
  12. Zhang, B. et al. (2022). “EntQA: Entity linking as question answering.” ICLR 2022.
  13. Orr, L. et al. (2021). “BOOTLEG: Chasing the tail with self-supervised named entity disambiguation.” VLDB 2021.
  14. Logeswaran, L. et al. (2019). “Zero-shot entity linking by reading entity descriptions.” ACL 2019, 3449–3460.
  15. Gillick, D. et al. (2019). “Learning dense representations for entity retrieval.” CoNLL 2019, 528–537.
  16. Rao, D. et al. (2013). “Entity linking: Finding extracted entities in a knowledge base.” Multi-source, Multilingual Information Extraction and Summarisation, Springer.
  17. Bunescu, R. & Pasca, M. (2006). “Using encyclopedic knowledge for named entity disambiguation.” EACL 2006, 9–16.
  18. Hoffart, J. et al. (2011). “Robust disambiguation of named entities in text (AIDA).” EMNLP 2011, 782–792.
  19. Kolitsas, N., Ganea, O. E. & Hofmann, T. (2018). “End-to-end neural entity linking.” CoNLL 2018, 519–529.
  20. Upadhyay, S. et al. (2018). “Joint multilingual supervision for cross-lingual entity linking.” EMNLP 2018, 2486–2495.
  21. Chen, W. et al. (2021). “Improving entity linking by modeling latent entity type information.” AAAI 2021, 12529–12537.
  22. OncsNET Team (2024). “OneNet: A fine-tuning free framework for few-shot entity linking via large language model prompting.” EMNLP 2024 Main. https://aclanthology.org/2024.emnlp-main.756/
  23. SynCABEL Team (2025). “Synthetic contextualized augmentation for biomedical entity linking.” arXiv:2601.19667. https://arxiv.org/html/2601.19667
  24. EntGPT Team (2024). “EntGPT: Entity linking with generative large language models.” arXiv:2402.06738. https://arxiv.org/html/2402.06738v2
  25. Learning from Negative Samples Team (2024). “Learning from negative samples in biomedical generative entity linking.” arXiv:2408.16493. https://arxiv.org/pdf/2408.16493
  26. Guiding LLMs Team (2025). “Guiding large language models for biomedical entity linking.” EMNLP Findings 2025. https://aclanthology.org/2025.findings-emnlp.1292.pdf
  27. Hybrid X-Linker Team (2024). “Hybrid X-Linker: Automated data generation and extreme multi-label ranking for biomedical entity linking.” arXiv:2407.06292. https://arxiv.org/pdf/2407.06292
  28. Shen, Z. et al. (2023). “A fair and in-depth evaluation of existing end-to-end entity linking systems.” arXiv:2305.14937. https://arxiv.org/pdf/2305.14937

  Key Terminology

• Surface Form: The literal text string spanning an entity mention in a document, e.g. “Apple” in “Apple announced new products”. The same entity may have many surface forms; the same surface form may refer to many entities.

• Candidate Generation: The first-pass retrieval of plausible entity referents from the knowledge base for a given surface form, using alias dictionaries and/or dense Embedding retrieval. Bounds the recall of the downstream disambiguation step.

• Entity Disambiguation (ED): The ranking or selection step that picks the correct entity from the candidate set generated for a given mention. Can use local context, entity description similarity, or global coherence signals.

• NIL: The decision that a mention has no referent in the target knowledge base; critical for open-world EL. Precision and recall on NIL decisions are tracked separately in TAC-KBP benchmarks.

• Bi-Encoder: A dual-tower Transformer Architecture that independently encodes mentions and entities into a shared dense vector space for efficient approximate nearest-neighbour retrieval. Faster than cross-encoders but less accurate on fine-grained disambiguation.

• Cross-Encoder: A Transformer Architecture that jointly encodes a mention-entity pair with full cross-attention, producing a fine-grained compatibility score. More accurate than bi-encoders but cannot scale to full-KB retrieval; typically applied to top-K bi-encoder candidates.

• Wikification: Entity linking specifically to Wikipedia articles or Wikidata QIDs, popularised by the TAC KBP track and by Google’s Knowledge Graph construction pipeline.

• Trie Decoding (Constrained Decoding): Generative decoding constrained by a prefix tree data structure to ensure outputs are valid entity names or identifiers. Used in GENRE and mGENRE to prevent hallucination of non-existent entities.

• Zero-Shot EL: Entity linking to entities unseen during training, using only entity description text as the linking signal; critical for fast domain adaptation to new ontologies or recently created entities.

• End-to-End EL: Systems that jointly perform mention detection and entity disambiguation in a single model pass, without a separate Named Entity Recognition step; eliminates error propagation between pipeline stages.

• Prior Probability P(entity | surface form): The probability that a given surface form refers to a given entity, estimated from anchor text frequency in Wikipedia. A simple but strong baseline signal for disambiguation; “Apple” most often refers to the technology company in a Wikipedia-derived prior.

• Global Coherence: The property of a collective entity assignment where all linked entities in a document are mutually compatible — consistent with the topic of the document and non-contradictory. Modelled via entity-entity relatedness scores or graph-based inference.

Provenance