Span-Level Classification

Unlike token-level approaches that assign labels to individual words, span-level classification identifies the exact start and end of a negated or uncertain phrase. This prevents fragmented outputs where 'no evidence of acute pulmonary embolism' might incorrectly label only 'pulmonary' as negated while missing 'embolism'. The model learns to predict span boundaries jointly, ensuring that multi-word clinical concepts like 'myocardial infarction' or 'acute respiratory distress syndrome' are classified as complete, coherent units.

The system first enumerates all possible token spans up to a maximum length within a sentence, then scores each candidate for negation or uncertainty. This exhaustive approach ensures no clinical entity is missed due to greedy decoding. For a sentence of length n, the model evaluates O(n²) spans, applying learned representations to each. While computationally intensive, modern GPU batching makes this tractable for clinical text, which typically contains short, focused sentences in radiology and pathology reports.

Span-level models simultaneously predict negation scope and polarity in a single forward pass. Rather than piping cue detection into a separate scope resolution step, the architecture learns the interaction between triggers like 'denies' and their syntactic dependents end-to-end. This joint modeling captures long-distance dependencies—for instance, correctly associating 'no evidence of' with a finding 15 tokens away—without relying on brittle rule-based windowing heuristics that fail on complex clinical syntax.

Each candidate span is encoded using its boundary tokens and internal attention-weighted sum from a transformer backbone like BioBERT or PubMedBERT. The representation fuses: (1) the hidden states of the span's start and end tokens, (2) a soft attention-weighted average of all tokens within the span, and (3) a learned width embedding encoding span length. This rich representation allows the classifier to distinguish 'no chest pain' (negated) from 'chest pain' (affirmed) based on subtle contextual cues that token-level CRF models miss.

Span-level models learn to distinguish true negation from pseudo-negation constructions where trigger words do not invert meaning. Examples include 'not only pneumonia but also sepsis' or 'did not rule out infection'. Because the model evaluates the entire span context—including words following the apparent trigger—it can learn that 'not only...but also' patterns are affirmative. This contextual disambiguation dramatically reduces false positives that plague regex-based systems, which mechanically flag any sentence containing 'no' or 'not' within a fixed window.

The contrasting approach where each individual token receives its own label (e.g., B-NEG, I-NEG) using sequence labeling architectures like BiLSTM-CRF. Unlike span-level methods, token-level classification requires explicit boundary detection and can suffer from label inconsistency where adjacent tokens within the same negation scope receive conflicting predictions. This approach is standard in benchmarks like the BioScope corpus but often requires post-processing to resolve fragmented spans.

The process of determining the exact contiguous sequence of tokens whose meaning is inverted by a negation cue. In span-level classification, the model directly predicts scope boundaries rather than labeling individual tokens. Key challenges include:

• Discontinuous scope: when a negation applies to multiple non-adjacent phrases
• Long-distance dependencies: where the cue and target are separated by many tokens
• Syntactic constraints: using dependency parse trees to define linguistically valid scope boundaries

The broader clinical NLP task of assigning a factuality label to each medical concept, categorizing it as present, absent, uncertain, or historical. Span-level classification directly outputs these assertion labels for entire concept spans rather than individual tokens. This aligns with the 2010 i2b2/VA challenge format, where systems must classify the assertion status of pre-identified clinical entities. Span-level methods naturally handle multi-word concepts like 'acute myocardial infarction' as single classification units.