Prompt-Based Disambiguation

The core mechanism involves constructing a prompt with a [MASK] token. The model predicts the most probable token to fill the mask, effectively performing disambiguation.

• Template: 'The patient's MI refers to [MASK].'
• Prediction: The model scores candidate tokens like 'infarction' or 'insufficiency' based on contextual probability.
• Key Advantage: Leverages the model's pre-trained objective directly, requiring no additional classification heads.

For autoregressive models, the task is framed as a direct instruction. The model generates the full expansion rather than selecting from a fixed inventory.

• Prompt: 'Expand the medical abbreviation MI in the following sentence: The patient was admitted with an acute MI.'
• Output: 'Myocardial Infarction'
• Constraint: Requires careful output parsing and validation against a known sense inventory like UMLS to prevent hallucinated expansions.

Candidate meanings are injected directly into the prompt to constrain the model's output space, transforming an open-generation problem into a multiple-choice classification.

• Prompt: 'The patient's MI refers to: A) Myocardial Infarction, B) Mitral Insufficiency, C) Mental Illness. Answer:'
• Mechanism: The model scores each option by calculating the probability of generating the corresponding label ('A', 'B', or 'C').
• Benefit: Eliminates the risk of generating a valid-sounding but incorrect expansion not found in the target terminology.

Disambiguation accuracy is improved by providing a small number of resolved examples within the prompt before the target instance, teaching the model the task pattern without weight updates.

• Example 1: 'Context: The patient has a history of CHF. Abbreviation: CHF. Expansion: Congestive Heart Failure.'
• Example 2: 'Context: The biopsy showed BCC. Abbreviation: BCC. Expansion: Basal Cell Carcinoma.'
• Target: 'Context: The echo showed severe MR. Abbreviation: MR. Expansion:'
• Impact: Dramatically improves performance on rare senses by demonstrating the expected mapping behavior.

Prompt-based methods provide a native, interpretable confidence metric by extracting the model's predicted probability for the chosen token or sequence.

• Calculation: Apply a softmax function to the logits at the [MASK] position or sum the log-probabilities of generated tokens.
• Thresholding: Predictions with a probability below a set cosine similarity threshold (e.g., < 0.7) can be flagged for human review.
• Clinical Utility: This directly supports human-in-the-loop review interfaces by prioritizing low-confidence disambiguations for a clinical documentation integrity specialist.

The prompt is enriched with metadata about the document's structure to resolve locally ambiguous abbreviations. The model uses section headers as a strong prior signal.

• Prompt: 'Section: Past Medical History. Text: MI in 2015. The abbreviation MI refers to [MASK].'
• Prompt: 'Section: Physical Exam. Text: No MI appreciated. The abbreviation MI refers to [MASK].'
• Outcome: The model learns to associate 'Past Medical History' with 'Myocardial Infarction' and 'Physical Exam' with 'Mitral Insufficiency,' demonstrating section header awareness.

The foundational computational task of identifying which meaning of a polysemous word is activated by its context. In clinical NLP, WSD resolves whether MI refers to Myocardial Infarction, Mitral Insufficiency, or Mental Illness based on surrounding text. Prompt-based methods reframe WSD as a masked language modeling problem, where the model predicts the correct sense from a structured inventory like UMLS.

A dynamic vector representation where a word's meaning shifts based on surrounding text. Unlike static embeddings, models like ClinicalBERT generate distinct vectors for 'MI' in 'The patient's MI was treated with tPA' versus 'The patient's MI was referred to dermatology.' Prompt-based disambiguation leverages these embeddings by comparing the masked token representation against candidate sense embeddings using cosine similarity.

The initial retrieval step that gathers all possible meanings of an abbreviation from a pre-compiled sense inventory. For 'CHF,' this might return:

• Congestive Heart Failure (C0018802)
• Chick Embryo Fibroblast (C0008138) The prompt then presents these candidates to the model for scoring, often constrained by semantic type filtering to eliminate implausible senses based on document context.

A mechanism in transformer architectures that allows the model to weigh the importance of different context words when resolving ambiguity. For 'The patient's MI was treated with tPA,' attention heads assign high weight to 'tPA' (a thrombolytic), signaling a cardiovascular context. Prompt-based methods exploit attention patterns by structuring prompts that guide the model to attend to clinically relevant surrounding tokens.

A disambiguation technique that constrains candidate meanings based on high-level UMLS categories. When resolving an ambiguous acronym, the system filters candidates by semantic type:

• T047 (Disease or Syndrome)
• T121 (Pharmacologic Substance)
• T059 (Laboratory Procedure) This dramatically reduces the candidate space and prevents cross-domain errors like interpreting a medication abbreviation as a procedure.

A machine learning paradigm where a model resolves abbreviation ambiguity using only a small number of labeled examples per sense. Prompt-based methods excel here by providing in-context demonstrations within the prompt itself. For example, showing the model one resolved instance of 'MI' as 'Myocardial Infarction' before asking it to disambiguate a new occurrence. This approach minimizes the need for large annotated clinical corpora.