Zero-Shot Entity Linking

Encodes textual mentions into a shared semantic vector space that is independent of the source language. This allows a mention in Finnish ('Yhdistyneet kansakunnat') and its English equivalent ('United Nations') to map to nearly identical vector coordinates.

• Uses models like LaBSE or XLM-RoBERTa as the encoder backbone
• Eliminates the need for language-specific feature engineering
• Enables direct cosine similarity comparison against a knowledge base index

Retrieves a shortlist of candidate entities from a multilingual knowledge base index using approximate nearest neighbor (ANN) search. The system queries a vector store containing entity descriptions encoded in multiple languages.

• Leverages FAISS or ScaNN for low-latency retrieval
• Candidate sets are typically limited to the top-50 or top-100 entities
• Works even when the knowledge base has no entries in the source language

A cross-encoder re-ranker scores each (mention, candidate entity) pair with full cross-attention. The model evaluates the semantic compatibility between the mention's context and the entity's description, even when they are in different scripts.

• Provides higher precision than bi-encoder similarity alone
• Handles code-switching and mixed-script mentions gracefully
• Trained on high-resource languages but generalizes to unseen ones

Applies Unicode normalization (NFC/NFD) and optional transliteration to handle surface-form mismatches. A mention in Cyrillic script can be matched to a Latin-script knowledge base entry by mapping both to a shared phonological representation.

• Mitigates lexical sparsity in morphologically rich languages
• Uses rule-based and learned transliteration models
• Critical for languages like Japanese, Arabic, and Hindi

Addresses the hubness problem in high-dimensional cross-lingual spaces, where certain entity vectors become universal nearest neighbors and degrade retrieval accuracy. Techniques like inverted softmax and cross-domain similarity local scaling (CSLS) correct this distortion.

• Prevents a small set of 'hub' entities from dominating all queries
• Improves recall@1 by up to 15 percentage points in low-resource settings
• Applied as a post-processing step on the embedding index

Identifies when a textual mention has no corresponding entity in the target knowledge base. The system learns a confidence threshold on the cross-encoder score, below which it predicts NIL rather than forcing an incorrect link.

• Prevents hallucinated entity links for out-of-KB concepts
• Uses a calibrated softmax temperature on the disambiguation head
• Essential for production systems dealing with emerging entities and rare long-tail concepts

The full sequence of grounding a textual mention to a knowledge base entry involves three stages:

• Mention Detection: Identifying the text span that refers to an entity
• Candidate Generation: Retrieving a shortlist of possible KB entries using alias tables or dense retrieval
• Entity Disambiguation: Ranking candidates by contextual compatibility In zero-shot scenarios, the disambiguation step must rely purely on cross-lingual semantic similarity rather than language-specific alias dictionaries.

The target of entity linking is a structured knowledge base such as Wikidata, DBpedia, or a proprietary enterprise graph. Each entity is identified by a unique URI (e.g., wd:Q567) and described by multilingual labels, aliases, and properties. Zero-shot linking exploits the fact that these KBs often contain language-agnostic identifiers and rich structural relationships that transcend any single language's lexical surface forms.

Zero-shot entity linking is a specific instance of cross-lingual transfer learning. A model is fine-tuned on a high-resource source language (typically English) with abundant labeled entity linking data. At inference time, the model's language-agnostic encoder projects a mention in an unseen target language into the same vector space, where it can be matched against KB entity embeddings. Success depends on the encoder's ability to produce isomorphic representations across languages.

Key datasets for measuring zero-shot entity linking performance include:

• Mewsli-9: Covers 9 languages with entity mentions linked to Wikidata
• TR2016: A cross-lingual entity linking dataset for Tweets
• AIDA CoNLL-YAGO: The standard monolingual benchmark adapted for cross-lingual evaluation Metrics focus on precision@k and mean reciprocal rank (MRR), measuring whether the correct KB entity appears in the top retrieved candidates.

The candidate generation stage in zero-shot linking increasingly uses multilingual dense passage retrieval (mDPR). Instead of relying on language-specific alias tables, a dual-encoder architecture encodes the mention and all KB entity descriptions into a shared dense vector space. This allows the retriever to fetch relevant candidates purely through semantic similarity, bypassing the need for translation or transliteration of the mention text.