Polarity Classification

While basic systems perform binary classification (positive/negative), production clinical NLP requires multi-class categorization to capture the full spectrum of factuality.

• Affirmed: The condition is definitively present ("patient has pneumonia")
• Negated: The condition is ruled out ("no evidence of pneumonia")
• Uncertain: The diagnosis is speculative ("possible pneumonia")
• Historical: Condition occurred in the past ("history of pneumonia")
• Experiencer: Condition applies to someone other than the patient ("mother had pneumonia")

The ConText algorithm extended polarity beyond simple binary negation to include these contextual modifiers, dramatically reducing false positive extractions.

Polarity classification depends on identifying negation cues — lexical triggers that invert meaning — and determining their scope, the span of text they modify.

Common clinical negation triggers include:

• Explicit negators: "no," "denies," "without," "absent"
• Pattern-based triggers: "negative for," "ruled out," "unremarkable"
• Pseudo-negation traps: "not only pneumonia but also" (does not negate)

The Negex algorithm pioneered rule-based scope determination using regular expressions and trigger-to-target distance thresholds. Modern transformer models like NegBERT learn scope boundaries from contextual embeddings, achieving higher accuracy on complex syntactic structures.

A critical challenge in polarity classification is distinguishing true negation from pseudo-negation — constructions containing negation words that do not actually reverse clinical meaning.

Examples of pseudo-negation patterns:

• Double negation: "not unlikely" (affirmative meaning)
• Emphatic affirmation: "not only diabetes but also hypertension" (both present)
• Rhetorical negation: "why not treat with antibiotics" (suggestion, not negation)

Failure to disambiguate these patterns causes false positive negation, where conditions are incorrectly marked as absent. Advanced systems use contextual embeddings to distinguish these cases by modeling the semantic intent of the full phrase rather than relying on trigger word presence alone.

Modern polarity classifiers output probabilistic confidence scores alongside categorical labels, enabling downstream systems to implement risk-based handling strategies.

Key applications of confidence scoring:

• High-confidence negations (>0.95): Automatically exclude from structured problem lists
• Borderline predictions (0.50-0.95): Route to human-in-the-loop review queues
• Low-confidence results (<0.50): Flag for clinical validation before any downstream use

This graduated approach prevents the two most dangerous failure modes: false negatives (missing a negation and incorrectly attributing a disease to a patient) and false positives (over-negating and suppressing a genuine diagnosis). Production systems typically optimize for negation precision to minimize false alarms.

Polarity classification extends beyond simple presence/absence to capture temporal context and experiencer attribution, preventing critical data contamination.

Temporal negation distinguishes:

• "Patient denies chest pain" (currently absent)
• "Chest pain resolved yesterday" (negated in present, affirmed in past)
• "No longer experiencing palpitations" (historical condition)

Experiencer negation prevents attribution errors:

• "Father died of myocardial infarction at 62" (condition applies to family member, not patient)
• "Sister has diabetes" (should not populate patient's problem list)

The ConText algorithm explicitly models these dimensions, ensuring that family history and resolved conditions are correctly classified rather than being conflated with current affirmed findings.

Polarity classification in healthcare demands rigorous evaluation beyond standard accuracy, with metrics weighted toward clinical safety.

Critical performance indicators:

• Negation Recall (Sensitivity): Proportion of actual negated findings correctly detected. Low recall means diseases are falsely attributed to patients — a direct patient safety risk
• Negation Precision: Proportion of flagged negations that are truly negated. Low precision creates alert fatigue from false alarms
• F1-Score by Class: Per-class performance reveals whether the model systematically fails on specific factuality categories
• Scope-Level Accuracy: Evaluates whether the exact token span of negation is correctly identified, not just the sentence-level label

The BioScope corpus provides gold-standard annotations for benchmarking, containing clinical free-text, biological papers, and abstracts with token-level negation and speculation labels.