NegBERT

NegBERT performs span-level classification on the BioScope corpus, assigning labels to contiguous token sequences rather than isolated words. This approach captures the full negation scope—the exact span of text whose meaning is inverted by a negation cue like 'no evidence of'. Unlike binary sentence classifiers, token-level prediction enables precise identification of which specific clinical finding is negated within a complex sentence containing multiple medical concepts.

NegBERT generates contextual embeddings that dynamically represent words based on surrounding tokens, solving the pseudo-negation problem that plagues rule-based systems. For example, in 'not only pneumonia but also sepsis', the model recognizes that 'not' does not negate 'pneumonia' because the full phrase 'not only...but also' forms an affirmative construction. This eliminates false positives that Negex and ConText algorithms produce when encountering negation trigger words in non-negating contexts.

NegBERT simultaneously classifies both negation and speculation in a unified architecture, distinguishing between:

• Negated findings: 'The patient denies chest pain'
• Uncertain findings: 'Possible pulmonary embolism cannot be ruled out'
• Affirmed findings: 'The patient presents with acute dyspnea' This joint modeling captures the interaction between epistemic modality and negation, where uncertainty cues like 'suggestive of' modify the factuality status of clinical assertions without fully negating them.

NegBERT is fine-tuned on the BioScope Corpus, the standard benchmark containing:

• Clinical free-text radiology reports
• Biological full papers from PubMed
• Scientific abstracts with annotated negation and speculation cues This domain-specific adaptation enables the model to handle clinical shorthand, medical abbreviations, and complex syntactic structures that general-domain BERT models misinterpret. The corpus provides gold-standard labels for both negation cues and their scope tokens, enabling supervised learning of the full negation resolution pipeline.

Unlike unidirectional LSTM-based approaches, NegBERT's bidirectional transformer architecture processes each token in the context of both preceding and following words simultaneously. This is critical for resolving double negation patterns like 'not unlikely' where the interaction between two negation elements cancels out to form an affirmative statement. The model also correctly handles temporal negation ('no longer present') and experiencer negation ('patient's mother has diabetes') by attending to the full syntactic context surrounding each clinical entity.

NegBERT outputs probabilistic confidence scores for each negation and uncertainty prediction, enabling downstream human-in-the-loop review interfaces to threshold results. Predictions with confidence below a configurable threshold are flagged for clinical reviewer audit, ensuring that ambiguous cases—such as hedging phrases like 'cannot entirely exclude'—receive expert verification. This architecture directly addresses the false negative rate safety metric, where missed negation could incorrectly attribute a disease to a patient record.

The NLP task of classifying statements expressing doubt or speculation rather than confirmed fact. NegBERT was fine-tuned to detect both negation and speculation.

• Triggers: Hedging phrases like 'suggestive of', 'cannot rule out', or 'possible'.
• Linguistic Basis: Rooted in epistemic modality, the grammatical expression of a speaker's degree of commitment.
• Clinical Impact: Prevents tentative differential diagnoses from being extracted as confirmed active problems.

The foundational rule-based predecessors to NegBERT. Negex uses regular expressions to identify negation triggers and scope. ConText extends this to detect historical conditions, hypotheticals, and experiencer status.

• Strengths: High precision, transparent logic, no training data required.
• Limitations: Brittle to syntactic variation; struggles with complex double negation ('not unlikely').
• NegBERT Advantage: Contextual embeddings capture semantic meaning beyond simple lexical pattern matching.

The publicly available annotated dataset used to fine-tune NegBERT. It serves as the standard benchmark for negation and speculation detection.

• Composition: Clinical free-text, biological full papers, and abstracts.
• Annotations: Token-level labels for negation cues, speculation cues, and their scopes.
• Significance: Provides the ground truth for training models to distinguish affirmed, negated, and uncertain findings.

The output label assigned to a clinical entity indicating its factuality. NegBERT's token-level predictions are aggregated to determine this status.

• Classes: Present, Absent, Uncertain, Historical, or Associated with Someone Else (Experiencer).
• Temporal Nuance: Distinguishes temporal negation ('no longer present') from permanent absence.
• Downstream Use: Critical input for accurate patient timeline construction and clinical decision support.

Rigorous measurement of NegBERT's performance on factuality tasks.

• Negation Precision: The proportion of correctly identified negated findings out of all flagged negations. High precision prevents false alarms.
• False Negative Rate: The proportion of missed negated or uncertain findings. A critical safety metric—missing a negation can incorrectly attribute a disease to a patient.
• Confidence Scoring: Probabilistic outputs allowing downstream systems to threshold results or prioritize ambiguous cases for human-in-the-loop review.