Zero-Shot Entity Linking

Unlike traditional systems that memorize a fixed set of entity IDs, zero-shot models rely entirely on entity descriptions. The system encodes the textual definition of a candidate entity and compares it to the mention's context. This allows instant linking to newly added knowledge base entries without retraining, as the model learns to map semantic meaning rather than specific identifiers.

The architecture typically uses a dual-tower Bi-Encoder to independently encode the mention context and all candidate entity descriptions into a shared dense vector space. This enables sub-linear retrieval via FAISS indexing. The system computes a dot product between the mention vector and millions of entity vectors to generate a shortlist of candidates, scaling efficiently to billion-scale knowledge bases.

After fast retrieval, a Cross-Encoder reranks the top-K candidates with full cross-attention between the mention and each entity description. This step is computationally expensive but highly precise, as it models fine-grained lexical and semantic interactions. The zero-shot capability is strongest here, as the Cross-Encoder learns to detect textual entailment between a mention's context and a previously unseen entity's definition.

A robust zero-shot system must handle Out-of-KB (OOKB) entities. By setting a calibrated threshold on the final linking confidence score, the model can predict NIL when no candidate description sufficiently matches the mention. This prevents false positives against newly added entities that are superficially similar but semantically distinct, maintaining precision in dynamic knowledge bases.

Training relies on contrastive learning with hard negative mining. The model is shown positive pairs (mention, correct entity description) and negative pairs (mention, random or hard-mined entity descriptions). This forces the encoder to focus on discriminative semantic features rather than memorizing entity IDs, directly enabling the generalization to unseen entities at inference time.

Because the architecture decouples entity identity from language-specific surface forms, zero-shot models often exhibit strong cross-lingual transfer. A system trained on English entity descriptions can link a Spanish mention to the correct English knowledge base entry by comparing the multilingual semantic context against the canonical English description, eliminating the need for language-specific training data.

The foundational retrieval architecture that makes zero-shot linking computationally feasible. A Bi-Encoder independently encodes the mention context and the entity description into dense vectors, allowing pre-computation of all entity embeddings.

• Enables fast Maximum Inner Product Search (MIPS) over millions of candidates
• Zero-shot capability comes from encoding entity descriptions, not learned IDs
• Contrastive loss functions train the model to pull correct mention-entity pairs together in vector space

A high-precision second-stage model that scores the top-K candidates retrieved by the Bi-Encoder. Unlike the Bi-Encoder, it processes the concatenated mention and entity text through full cross-attention.

• Captures fine-grained lexical overlap and semantic interaction
• Too slow for full-scale retrieval, but essential for accurate final disambiguation
• In zero-shot settings, the Cross-Encoder must generalize its reasoning patterns to unfamiliar entity descriptions

The dense vector representation of all entities in a knowledge base, constructed from their textual descriptions. This is the key enabler of zero-shot transfer.

• Entities are represented by encoding their title, definition, and structured attributes
• New entities can be added to the index at any time without retraining
• The quality of the embedding space depends on the contrastive training objective and the richness of entity descriptions

The critical mechanism that allows a system to correctly output 'no match' when a mention refers to an entity absent from the knowledge base. Without robust NIL prediction, zero-shot systems hallucinate links.

• Typically implemented via a confidence threshold on the final linking score
• Advanced methods train a dedicated NIL classifier using features from both encoder stages
• Essential for production systems where knowledge bases are inherently incomplete

The dominant paradigm for creating large-scale training data without manual annotation. Entity mentions are heuristically aligned to knowledge base entries using anchor text links from sources like Wikipedia.

• Generates millions of weakly labeled mention-entity pairs
• Enables training on diverse entity types that support zero-shot generalization
• Requires careful noise filtering to avoid training on incorrect heuristic matches