📄 Speculation and negation annotation in natural language texts: what the case of BioScope might (not) reveal

Summary

The paper reflects on what building BioScope taught about annotating uncertainty and negation. Information extraction needs to keep factual statements apart from hedged or denied ones, and that requires suitably annotated corpora. BioScope is described as the largest resource with both negation and hedge cues and the only one that also marks linguistic scopes. It covers radiology reports, full biological papers and GENIA abstracts. Two linguistics students annotated it under the supervision of a linguist, who resolved disagreements. The author sets out the core guidelines: every cue has a scope that contains the cue, and cues are kept minimal while scopes are made maximal. She then goes through hard cases, including cues without a scope, ellipsis, multiword cues, overlapping and nested scopes, and ambiguous lexical items. BioScope is also compared with the event-centred GENIA Event corpus, and the two often disagree on what counts as uncertain. The lesson is that guidelines need to be fixed in advance but should keep evolving as the data reveals new cases.

Key points

• BioScope contains three biomedical text types: radiological reports, full papers and GENIA abstracts. A further 15 full papers with hedge annotation served as the CoNLL-2010 Shared Task evaluation set.
• BioScope uses a single speculation category, whereas GENIA Event splits uncertainty into doubtful and probable. BioScope also leaves weasel words unannotated, while WikiWeasel targets them.
• Scopes must cover a continuous span and may not intersect. When two scopes would overlap, one of them (usually the negative one) is widened. Nested scopes are read as stacked operators, as in "possibly, no such event…".
• In one example sentence GENIA marks only 2 of 6 events as probable, while all 6 fall inside BioScope's speculative scope.
• Negated investigation verbs in the present perfect (e.g. has not been described) are treated as doubtful in GENIA but as negative in BioScope.

Interesting ideas

• I1. Mark the cue as the smallest unit that expresses hedging or negation, and stretch its scope to the largest syntactic unit, which gives a clear and repeatable boundary rule.
• I2. A negative word can sit inside a hedge (cannot be excluded), so the mere presence of "not" does not make a sentence negated.
• I3. Whether an item is a cue can depend on its construction (indicate versus indicate that) or its sense (if).
• I4. An event sitting inside a speculative scope is not necessarily speculated itself. Only the asserted proposition carries the hedge, not every nominalised event embedded in it.
• I5. Comparing two annotation schemes on the same sentences exposes systematic disagreements, such as the present-perfect negation case.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): annotate each negation or hedge cue on its tokens and assign as scope the maximal UD subtree of the clause it governs, so every scope decision can be traced to the tree.
• I2 (frames): match multiword cue patterns before applying the single-word negation rule, so the phrase receives speculation rather than negation.
• I3 (tree): decide cue status from the dependency configuration, for example whether a ccomp clause is present, rather than from the word form alone.
• I4 (world): attach modality to the main asserted proposition and keep embedded events such as changes or transcription neutral, instead of marking everything in scope as uncertain.
• I5 (learn): before merging silver data drawn from different guidelines, diff them on a shared sample to find convention clashes rather than treating them as noise.