Candidate Sense Generation

The process of querying a pre-compiled knowledge base to retrieve all possible meanings for an ambiguous abbreviation. For clinical text, this inventory is typically derived from the Unified Medical Language System (UMLS) Metathesaurus, which aggregates over 200 source vocabularies.

• Input: The surface form 'MI'
• Output: A set of Concept Unique Identifiers (CUIs) including C0027051 (Myocardial Infarction) and C0026266 (Mitral Valve Insufficiency)
• Mechanism: Exact string matching against a normalized index of abbreviations and their long forms

A pruning step that constrains the candidate set based on high-level UMLS Semantic Types before contextual scoring begins. This eliminates grossly implausible senses and reduces computational load.

• A mention in a 'Medications' section can filter candidates to only those with semantic type 'Clinical Drug' or 'Pharmacologic Substance'
• Removes candidates like 'Disease or Syndrome' or 'Diagnostic Procedure' from consideration
• Reduces the candidate set by 40-60% in typical clinical narratives, improving downstream precision

The generation of additional surface forms and synonyms for each candidate sense to improve the robustness of subsequent similarity scoring. A single CUI may have dozens of lexical variants.

• C0027051 (Myocardial Infarction) expands to: 'MI', 'heart attack', 'myocardial infarct', 'cardiac infarction', 'coronary occlusion'
• Crucial for handling: Clinical shorthand, typographical errors, and institution-specific jargon
• Leverages the MRCONSO table in UMLS, which maps each concept to all known string forms across source vocabularies

An advanced candidate generation technique that traverses hierarchical relationships in the UMLS semantic network to include broader or narrower concepts as additional candidates.

• If 'ACE' maps to 'Angiotensin-Converting Enzyme Inhibitors', graph traversal can include child concepts like 'Lisinopril' and 'Enalapril'
• Uses PAR/CHD (parent/child) and RB/RN (broader/narrower) relationship types
• Enables the disambiguation system to resolve abbreviations to the correct level of granularity required by the downstream task, such as ICD-10-CM coding

A quantitative measure of how polysemous an abbreviation is, calculated as the cardinality of its candidate sense set. This index drives triage logic in the disambiguation pipeline.

• Low Ambiguity (1-2 senses): 'CHF' almost always maps to 'Congestive Heart Failure' — fast-track to expansion
• High Ambiguity (10+ senses): 'MI' requires full contextual scoring across cardiology, neurology, and dermatology domains
• Extreme Ambiguity (50+ senses): 'CA' (Cancer, Calcium, Cardiac Arrest, Carcinoma, etc.) triggers section-header awareness and document-level context integration

A proactive analysis that identifies the specific sense pairs most likely to be confused during scoring, based on co-occurrence statistics in the training corpus. This metadata is attached to the candidate set.

• Classic pair: 'MI' → Myocardial Infarction vs. Mitral Insufficiency (both cardiology, same semantic type)
• Metadata tag: Flags these candidates as requiring high-confidence contextual evidence for resolution
• Enables the downstream scoring model to apply contrastive attention mechanisms that focus on differentiating clinical features like 'troponin elevation' vs. 'regurgitant murmur'

The foundational sense inventory for candidate generation. UMLS integrates over 200 biomedical vocabularies, providing Concept Unique Identifiers (CUIs) and semantic types that serve as the target meanings for ambiguous abbreviations. A single abbreviation like 'CA' maps to multiple CUIs—'Carcinoma' (C0007097), 'Calcium' (C0006675), and 'California' (C1552032)—all retrieved during candidate generation.

A pruning technique applied immediately after candidate retrieval. By constraining candidates to high-level UMLS categories, the system eliminates implausible senses before scoring. For example, when 'MI' appears in a Medications section, candidates with semantic type 'Clinical Drug' are retained while 'Disease or Syndrome' senses are deprioritized. This reduces the search space and improves downstream accuracy.

The vector representation that enables candidate scoring. Models like ClinicalBERT generate dynamic embeddings where the same abbreviation receives different vectors depending on surrounding text. The embedding for 'CHF' in 'CHF exacerbation' is mathematically closer to the 'Congestive Heart Failure' sense embedding than to the 'Swiss Franc' sense, enabling precise disambiguation via cosine similarity.

The downstream task that consumes disambiguated candidates. Once the correct sense is selected, entity linking grounds the mention to a unique identifier in a knowledge base. Key targets include:

• SNOMED CT Concept IDs for clinical findings
• RxNorm CUIs (RxCUIs) for medications
• ICD-10-CM codes for billing classifications This step transforms resolved text into structured, queryable data.

An error analysis methodology for refining candidate generation. By identifying the specific sense pairs a model most frequently confuses—such as 'MI' for 'Myocardial Infarction' versus 'Mitral Insufficiency'—engineers can augment the sense inventory with discriminative features or add targeted training examples. This analysis directly informs improvements to the candidate retrieval and scoring pipeline.

A family of transformer models pre-trained on clinical corpora like MIMIC-III that provide the contextual representations for candidate scoring. Unlike general-domain BERT, ClinicalBERT captures domain-specific co-occurrence patterns—understanding that 'q.d.' near 'tablet' likely means 'once daily' rather than 'every other day'. These models form the backbone of modern disambiguation architectures.