BioScope Corpus

BioScope is not a single dataset but an assembly of three distinct text types, enabling evaluation across different medical writing styles:

• Clinical Free-Text: De-identified radiology reports and clinical documents containing high-density negation and hedging.
• Biological Full Papers: Full-text articles from the biological literature, representing formal scientific argumentation.
• Biological Abstracts: Concise summaries of scientific papers, testing performance on condensed, information-dense text. This diversity ensures models trained on BioScope generalize across both clinical and scientific domains.

The corpus uses a two-layer annotation scheme that captures the full semantics of factuality:

• Cue Annotation: Each word or phrase that triggers negation (e.g., 'no', 'denies') or uncertainty (e.g., 'possible', 'suggestive of') is tagged at the token level.
• Scope Annotation: For every cue, the contiguous span of text whose meaning is modified is marked. This teaches models the exact boundary of what is being negated or speculated. This dual-layer approach moves beyond simple keyword matching to true semantic understanding.

BioScope distinguishes between two primary factuality modifiers, enabling fine-grained classification:

• Negation: The explicit reversal of a clinical or biological finding's truth value. Example: 'The patient has no evidence of pneumothorax.'
• Speculation (Uncertainty): Expressions of doubt or hedging that weaken the commitment to a statement. Example: 'The mass is suspicious for malignancy.' This separation is critical because a negated finding is definitively absent, while a speculative finding requires further investigation.

The corpus captures challenging linguistic phenomena that simple keyword systems fail on, making it a robust testbed:

• Pseudo-Negation: Constructions like 'not only... but also' that contain negation words but do not negate.
• Double Negation: Phrases like 'not unlikely' that semantically cancel out to an affirmation.
• Historical Context: Findings negated in the present but affirmed in the past ('no longer present').
• Experiencer Shifts: Conditions that are affirmed but apply to a family member, not the patient.

Annotations are provided at the individual token level using BIO (Begin, Inside, Outside) tagging schemes. This granularity supports:

• Training sequence labeling models like BiLSTM-CRF and transformer-based token classifiers.
• Precise evaluation of scope boundary detection, not just cue identification.
• The development of systems that can pinpoint exactly which words in a clinical sentence are affected by a negation or uncertainty trigger.

The computational task of identifying linguistic cues that semantically reverse the existence of a clinical finding. Systems trained on BioScope learn to distinguish absent conditions from affirmed ones.

• Core Goal: Prevent false positive diagnoses in structured data.
• Key Triggers: 'no', 'denies', 'without evidence of'.
• Challenge: Distinguishing true negation from pseudo-negation ('not only pneumonia but...').

The NLP task of classifying statements expressing doubt or speculation regarding a medical condition. BioScope provides the gold-standard annotations for hedging and epistemic modality.

• Core Goal: Differentiate confirmed diagnoses from suspected ones.
• Key Triggers: 'possible', 'suggestive of', 'cannot rule out'.
• Clinical Impact: Prevents unconfirmed differential diagnoses from being recorded as active problems.

A widely adopted, rule-based regular expression algorithm that identifies negation triggers and their scope. It serves as the high-precision baseline against which BioScope-trained neural models are compared.

• Mechanism: Uses lexical patterns to invert the assertion status of clinical concepts.
• Limitation: Struggles with complex syntax and pseudo-negation that contextual models handle.

An extension of Negex that detects not only negation but also historical conditions, hypothetical statements, and the experiencer of a finding. BioScope's fine-grained annotations enable training models that match ConText's rule-based scope.

• Dimensions: Negation, Temporality, Experiencer.
• Example: 'Mother has diabetes' → Experiencer is family member, not patient.

A transformer-based language model specifically fine-tuned on the BioScope corpus for token-level negation and speculation detection. It leverages contextual embeddings to resolve semantic ambiguity.

• Architecture: BERT-base fine-tuned on BioScope abstracts and clinical text.
• Advantage: Captures long-range dependencies that rule-based systems miss.
• Output: Token-level labels for negation cues, scopes, and speculation.

The process of assigning a label to a clinical named entity indicating whether the concept is present, absent, or uncertain in the patient record. BioScope provides the training data for this core clinical NLP task.

• Categories: Affirmed, Negated, Historical, Hypothetical, Experiencer.
• Evaluation Metric: Precision on negated findings is critical to avoid attributing false conditions.