Lexical Variant Generation

The initial, high-recall retrieval stage where lexical variants are matched against a knowledge base index. This phase uses fast, approximate methods to fetch a small set of plausible entries for a given text mention.

• Lexical methods: BM25 and TF-IDF over generated morphological variants
• Dense methods: Approximate Nearest Neighbor (ANN) search over bi-encoder embeddings
• Hybrid approaches: Combining sparse lexical hits with dense semantic vectors

The quality of candidate generation directly depends on the comprehensiveness of the variant lookup table. Missing a valid orthographic form here means the entity can never be correctly linked downstream.

The second-stage process where a computationally intensive model scores and orders the candidate entities generated from lexical variants. Unlike the high-recall generation phase, ranking prioritizes precision.

• Cross-encoder rerankers process mention-candidate pairs jointly through a transformer
• Feature-based rankers combine semantic type, string similarity, and graph distance
• Confidence thresholding filters low-scoring candidates to prevent false positives

Effective variant generation reduces the ranking burden by ensuring the correct entity is present in the candidate set, allowing the ranker to focus on disambiguation rather than overcoming recall failures.

The process of resolving ambiguous clinical shorthand to its correct expanded form using contextual embeddings. Lexical variant generation must account for abbreviations as a distinct variant class.

• Context-dependent: 'CA' may expand to 'cancer', 'calcium', or 'cardiac arrest'
• Context-independent: 'MI' almost always expands to 'myocardial infarction' in cardiology notes
• Domain-specific lexicons: Specialty-dependent abbreviation maps improve expansion accuracy

Generating both the abbreviated form and all plausible expansions as variants ensures the entity linker can match regardless of which form appears in the target ontology.

The core challenge of resolving which specific concept an ambiguous term refers to based on surrounding clinical context. Lexical variants alone cannot solve this—they only ensure the correct concept is retrievable.

• Polysemy: 'Cold' could be a temperature sensation or a viral illness
• Homonymy: 'Lead' could be a metal, an ECG electrode, or a verb
• Contextual signals: Laterality, anatomy, temporality, and negation all constrain meaning

The variant generation system must produce forms for all possible senses, while the disambiguation model uses semantic type filtering and contextual cues to select the correct one.

The critical function of correctly identifying when a clinical mention has no corresponding concept in the target knowledge base. Over-generating variants can increase false-positive NIL predictions if not properly calibrated.

• Out-of-vocabulary terms: Novel drug names, rare diseases, or institution-specific jargon
• Misspellings: Variants that don't match any canonical form despite generation efforts
• Granularity mismatches: Mentions too specific or too general for the ontology

A robust variant generation pipeline includes confidence scoring on generated forms, allowing the linker to abstain rather than force an incorrect grounding when no variant matches with sufficient certainty.

A constraint applied during candidate retrieval that restricts potential matches to entities belonging to specific UMLS Semantic Types. This dramatically reduces the candidate space and improves precision.

• Type constraints: 'Disease or Syndrome' (T047), 'Pharmacologic Substance' (T121), 'Laboratory Procedure' (T059)
• Type coercion: Using context to infer the semantic type of an ambiguous mention
• Variant-type alignment: Ensuring generated variants are tagged with their expected semantic category

When generating lexical variants, preserving or inferring the semantic type of each variant form enables downstream filtering that eliminates cross-type false matches before ranking occurs.