Entity Linking

The initial stage of identifying spans of text that refer to entities. This involves named entity recognition (NER) to locate boundaries and classify the type (e.g., Person, Drug, Disease).

• Boundary Detection: Accurately capturing multi-word phrases like 'acute myocardial infarction'.
• Overlap Resolution: Handling nested or overlapping entity mentions.
• Ambiguity Flagging: Marking pronouns and acronyms for later resolution.

The process of retrieving a set of possible knowledge base identifiers for a detected mention. This narrows the search space from millions of concepts to a manageable few.

• Name Matching: Using exact string matching, fuzzy matching, or phonetic algorithms against a synonym dictionary.
• Acronym Expansion: Resolving 'MS' to 'Multiple Sclerosis' or 'Mitral Stenosis' based on context.
• Search Engine Indexing: Leveraging inverted indices built on the knowledge base for rapid lookups.

The core ranking step where the correct entity is selected from the candidate set by analyzing the surrounding text. This relies heavily on contextual embeddings from transformer models.

• Local Coherence: The chosen entity should be topically consistent with the surrounding sentence and paragraph.
• Global Coherence: All linked entities in a document should form a logically consistent semantic graph, maximizing collective agreement.
• Prior Probability: Incorporating the general popularity or commonness of an entity as a statistical baseline.

The critical capability to recognize when a valid mention does not have a corresponding entry in the target knowledge base. Linking to a 'NIL' identifier prevents false positive grounding.

• Confidence Thresholding: Setting a minimum similarity score below which a link is rejected.
• Out-of-Vocabulary Detection: Identifying emerging terms, rare diseases, or new drugs not yet cataloged in the ontology.
• Abstention Mechanisms: Training models to explicitly predict a 'no-link' class rather than forcing an incorrect match.

The backend infrastructure that stores the target ontology and enables sub-millisecond retrieval. Efficient indexing is essential for production systems processing high-volume clinical text.

• Alias Tables: Pre-computed mappings from every known surface form and synonym to its canonical ID.
• Dense Vector Search: Encoding entity descriptions into embeddings for semantic similarity search, catching paraphrases that string matching misses.
• Graph Traversal: Using the ontology's hierarchical structure (e.g., is_a relationships) to expand or refine candidate searches.

The extension of entity linking to connect mentions in one language to a knowledge base that may be primarily defined in another, such as linking a Spanish clinical note to English UMLS concepts.

• Translation-Based Methods: Using machine translation as a preprocessing step before monolingual linking.
• Joint Multilingual Embeddings: Training models on aligned corpora to map words and phrases from different languages into a shared semantic vector space.
• Language-Agnostic Knowledge Graphs: Utilizing multilingual ontologies that natively contain labels and synonyms in dozens of languages.

Also known as record linkage or deduplication. This process identifies when multiple records refer to the same real-world entity. In a healthcare context, this means merging duplicate patient records from different systems before linking clinical mentions.

• Uses blocking and scoring techniques to scale
• Critical for creating a clean master patient index
• Prevents fragmented knowledge graphs where a single patient is split across multiple nodes

The process of establishing logical mappings between disparate ontologies like SNOMED CT, ICD-10-CM, and LOINC. Entity Linking relies on these pre-aligned crosswalks to ground mentions to multiple target vocabularies simultaneously.

• Uses lexical matching, structural graph matching, and logical reasoning
• Enables a single annotation to satisfy billing, clinical, and research use cases
• Maintained through the Observable Medical Ontology System (OMOS)

The upstream task that identifies semantic relationships between entities. Once entities are linked to a knowledge base, Relation Extraction populates the edges of the graph by classifying connections like 'treats', 'causes', or 'located_in'.

• Often performed jointly with Entity Linking in end-to-end neural models
• Transforms a bag of linked entities into a connected clinical narrative
• Uses attention mechanisms to find syntactic paths between entity pairs

The predictive task of inferring missing links in a graph. After Entity Linking grounds mentions to existing nodes, link prediction models can suggest new relationships that were never explicitly stated in text.

• Uses translational distance models like TransE or tensor factorization
• Enhances clinical decision support by surfacing latent drug-disease associations
• Validated against held-out triples during training

Graph Retrieval-Augmented Generation uses a knowledge graph as the retrieval source for large language models. Entity Linking is the ingestion pipeline that populates this graph from unstructured documents, enabling the LLM to traverse structured relationships.

• Grounds LLM responses in deterministic, curated facts
• Reduces hallucination by constraining generation to graph paths
• Combines vector similarity search with graph traversal for multi-hop reasoning