Few-Shot Disambiguation

The disambiguation task is reframed as a text-to-text generation problem using structured prompts. Instead of fine-tuning a classifier head, the model is given a template like 'The abbreviation MI in the context of cardiology refers to [MASK].'

• Mechanism: Leverages the pre-trained knowledge of large language models.
• Advantage: Eliminates the need for a large, task-specific labeled dataset.
• Example: A prompt can explicitly ask the model to choose between 'Myocardial Infarction' and 'Mitral Insufficiency' based on the provided sentence.

This method generates a contextual embedding for the ambiguous abbreviation and compares it against pre-computed embeddings of all possible candidate senses using cosine similarity.

• Process: The candidate sense with the highest similarity score is selected.
• Key Tech: Relies on models like ClinicalBERT to produce dynamic, context-aware vectors.
• Efficiency: Only a few labeled examples are needed to calibrate the similarity threshold.

A few hand-picked examples demonstrating correct disambiguation are inserted directly into the model's context window before the query. The model learns the pattern from these exemplars without any weight updates.

• Format: A series of Input: '...' Output: '...' pairs followed by the target input.
• Selection: Exemplar selection is critical; choosing examples that are semantically close to the query improves accuracy.
• Benefit: Allows for rapid adaptation to new abbreviation types without retraining.

The initial step retrieves all possible meanings of an abbreviation from a sense inventory like the Unified Medical Language System (UMLS). For 'CHF', this would include 'Congestive Heart Failure' and 'Chick Heart Fibroblast'.

• Source: UMLS Metathesaurus, SNOMED CT, or custom institutional lexicons.
• Filtering: Semantic type filtering can immediately prune candidates (e.g., keep only 'Disease or Syndrome' types).
• Role: Creates the constrained choice set from which the few-shot model selects.

An error analysis technique that identifies the specific sense pairs a model most frequently confuses, such as 'MI' for 'Myocardial Infarction' vs. 'Mitral Insufficiency'.

• Method: A confusion matrix is generated from a small validation set.
• Action: Targeted few-shot examples are then crafted specifically to disambiguate these high-confusion pairs.
• Outcome: Drastically improves model precision on clinically dangerous ambiguities with minimal data.

A model's ability to use the title of a clinical document section as a strong prior signal. An abbreviation found under 'Past Medical History' is disambiguated differently than one under 'Family History'.

• Signal: The section header is explicitly included in the prompt or as a special token.
• Impact: Provides a document-level context that often resolves ambiguity instantly.
• Few-Shot Link: A single example showing how a section header changes the meaning can teach the model this behavior.

A dynamic vector representation where a word's meaning shifts based on surrounding text. Unlike static embeddings (Word2Vec), transformer models like BERT generate distinct vectors for 'MI' in 'The patient's MI was treated with thrombolytics' versus 'The MI showed a suspicious lesion.' This context-sensitivity is the foundational mechanism enabling few-shot disambiguation—the model learns to separate senses by observing how the same surface form appears in different distributional contexts, requiring only a handful of examples per sense to establish decision boundaries.

A technique that reframes abbreviation resolution as a masked language modeling or generative task. Instead of training a classifier head, the model receives a structured prompt like 'The patient's MI refers to [MASK]' and predicts the correct expansion from its pre-trained knowledge. This approach leverages the model's existing semantic understanding, making it highly effective in few-shot regimes. Instruction-tuned models can perform zero-shot disambiguation when prompted with a sense inventory and explicit formatting constraints, eliminating the need for task-specific fine-tuning entirely.

The UMLS Metathesaurus serves as the definitive sense inventory for clinical disambiguation, containing millions of concepts with unique identifiers (CUIs). When a model encounters 'CHF,' it queries the UMLS to retrieve candidate senses: C0018802 (Congestive Heart Failure) and C0008073 (Chick Embryo Fibroblast). Each CUI carries semantic types—'Disease or Syndrome' versus 'Cell'—providing high-level constraints. Few-shot disambiguation systems use UMLS concept embeddings pre-computed from definitional text to score candidate senses against the contextualized mention embedding via cosine similarity.

A constraint-based technique that narrows candidate senses by their UMLS Semantic Network categories before scoring. When resolving 'MI,' the system filters candidates to only those with semantic types relevant to the clinical context—'Disease or Syndrome' (Myocardial Infarction) and 'Anatomical Abnormality' (Mitral Insufficiency)—while excluding irrelevant types like 'Laboratory Procedure.' This dramatically reduces the candidate space and prevents absurd mappings. In few-shot settings, semantic type constraints act as a strong inductive bias, allowing models to generalize from minimal examples by ruling out entire categories of incorrect senses.

An error analysis methodology that identifies the specific sense pairs a disambiguation model most frequently confuses—such as 'MI' for Myocardial Infarction versus Mitral Insufficiency. By aggregating prediction errors on a validation set, engineers discover that these two cardiovascular senses share overlapping context words ('chest pain,' 'cardiac'), making them harder to separate than 'MI' versus 'Michigan.' In few-shot learning, confusion pair analysis is critical: it reveals which senses require additional carefully-chosen examples to establish clear decision boundaries, guiding active learning sample selection for maximum labeling efficiency.

The use of information beyond the immediate sentence to resolve local ambiguity. A mention of 'MS' in a single sentence might be ambiguous, but the problem list containing 'Multiple Sclerosis' or the section header 'Neurology Consult' provides strong disambiguating signals. Few-shot systems exploit document structure—SOAP note sections, ICD-10 codes in the encounter header, medication lists—as additional features without requiring explicit annotation. This distant supervision from document metadata effectively multiplies the information available from each labeled example, making few-shot learning viable for complex clinical narratives.