Candidate Ranking

The dominant architecture for candidate ranking that processes the mention-candidate pair jointly through a transformer model. Unlike bi-encoders that encode mentions and entities separately, cross-encoders apply full self-attention across the concatenated sequence, allowing the model to capture subtle lexical overlaps, semantic contradictions, and contextual nuances. This joint processing produces a high-fidelity relevance score but is computationally expensive, making it suitable only for reranking a small candidate set rather than full-scale retrieval.

A training paradigm where the model learns to maximize the margin between the score of the correct entity and all incorrect candidates. By using hard negative mining, the system selects confusable entities—such as 'Diabetes Mellitus' vs 'Diabetes Insipidus'—and forces the model to distinguish them. The output is a calibrated probability distribution over the candidate set, enabling downstream confidence thresholding for human-in-the-loop review.

The critical function of identifying when no valid match exists in the target knowledge base. The ranking model must output a confidence score below a calibrated threshold to trigger a NIL prediction, preventing false grounding of novel or rare clinical concepts. This is often implemented as an additional 'no-match' embedding in the candidate set, trained against mentions known to lack corresponding UMLS CUIs.

The ranking model encodes the target mention within its surrounding sentence or paragraph window to resolve ambiguity. For instance, the mention 'cold' in 'patient complains of cold extremities' vs 'patient has a cold' requires the model to attend to neighboring tokens. Transformer-based rankers use positional embeddings and self-attention to weight the contextual relevance of each token, ensuring the disambiguation is evidence-based.

A ranking signal that computes the semantic similarity between the clinical mention context and the textual definition or synonyms of each candidate entity. By encoding entity descriptions from sources like SNOMED CT fully specified names or UMLS definitions, the model can match mentions to concepts even when exact lexical overlap is absent. This is particularly effective for linking colloquial clinical shorthand to formal ontological terms.

A neural architecture that processes a mention-candidate pair jointly through a transformer to produce a high-fidelity relevance score. Unlike bi-encoders, cross-encoders attend to the full interaction between the clinical mention and the candidate entity definition, enabling precise disambiguation.

• Mechanism: Concatenates mention text and entity description, passes through BERT-style model, outputs single logit
• Trade-off: Highest accuracy but computationally expensive; used only on top-k candidates
• Example: A cross-encoder distinguishes 'MI' as Myocardial Infarction vs. Mitral Insufficiency by attending to chest pain context

A dual-tower neural network that independently encodes a text mention and a knowledge base entity into dense vectors for efficient semantic similarity search. Bi-encoders enable fast candidate retrieval before the ranking stage.

• Mention Tower: Encodes the clinical text span and its surrounding context
• Entity Tower: Encodes the canonical entity description from the knowledge base
• Similarity: Cosine similarity between the two output vectors
• Advantage: Entity embeddings can be pre-computed and indexed for sub-linear retrieval

The process of adjusting a model's predicted probability for a linked entity to ensure it accurately reflects the true likelihood of correctness. Raw softmax scores from neural rankers are often overconfident.

• Platt Scaling: Fits a logistic regression on validation set logits
• Isotonic Regression: Learns a non-decreasing calibration function
• Temperature Scaling: Divides logits by a learned temperature parameter
• Clinical Impact: Proper calibration enables reliable NIL prediction and human-in-the-loop review thresholds

A contrastive learning strategy that selects highly confusable but incorrect candidate entities during training to improve a model's disambiguation capability. Standard random negatives are too easy and don't teach fine-grained distinctions.

• Method: For each mention, retrieve top-k candidates using a bi-encoder, then treat all non-gold candidates as hard negatives
• Example: For 'COPD exacerbation', hard negatives include 'Asthma exacerbation' and 'Bronchiectasis flare'
• Benefit: Forces the ranker to learn subtle semantic boundaries between closely related clinical concepts

The critical entity linking function of correctly identifying when a clinical mention has no corresponding concept in the target knowledge base, preventing false grounding. This is essential for real-world clinical text where novel or non-standard terms appear.

• Threshold Method: If the top candidate's confidence score falls below a calibrated threshold, predict NIL
• Learned NIL: Train a dedicated 'no-match' embedding vector that competes with real candidates
• Clinical Risk: False grounding of an unrecognized drug or device can propagate errors into downstream decision support