📄 Combining Manual Rules and Supervised Learning for Hedge Cue and Scope Detection

Summary

The paper tackles hedging in biomedical writing: finding the words that weaken a claim (cues) and the stretch of the sentence they weaken (scope). It was built for the CoNLL-2010 shared task, trained on a revised BioScope corpus (9 full papers and 1273 abstracts) and tested on a new set of full papers. Every sentence is parsed with RASP. Cue detection uses a CRF tagger with FILO labels and features drawn from RASP grammatical relations (GRs), lemmas and POS. Its output is then checked against the cues seen in training, and anything unmatched is dropped. For scope, hand-written rules walk the GR graph, and the rule chosen depends on the cue's POS class. A second CRF corrects the rule output, and a few post-processing rules turn the token tags into continuous scopes. With predicted cues, scope F-score climbs from 0.50 (baseline) to 0.53 (rules) to about 0.56 (rules plus CRF). The authors also argue that scopes defined over the surface string are a poor fit for what hedging really affects.

Key points

• Training data: 2670 sentences from papers and 11,871 from abstracts, with 682 and 2694 cues respectively. Most scopes begin with their cue (72% in papers, about 81% in abstracts).
• Lemma is the most useful feature for cue detection. Unseen cues are handled poorly, so the system keeps only predictions that match a cue from training, using the longest match. This raises precision without hurting recall.
• There are eight scope rules, selected by the cue's generalised POS: modal VM, verb VV, adjective JJ, adverb RR, noun NN, conjunction CC, "whether" CSW, and a default. Graph traversal halts at text-adjunct GRs and clause separators. Passive constructions and raising verbs such as appear/seem trigger extra rules that pull in the subject.
• Results with predicted cues on 1033 evaluation scopes: baseline F 0.500, manual rules 0.525, rules plus second CRF 0.556 to 0.559. With gold cues: baseline 0.57, rules 0.64.
• Error analysis of the rules: about 35% of wrong scopes come from wrong or fragmented RASP graphs, and about 65% from cases the rules do not cover.

Interesting ideas

• I1. Hand-written rules make a first guess, and their output is then given to a learned sequence tagger as one more feature, so the learner only has to fix the exceptions instead of learning the whole pattern.
• I2. A learned detector's predictions are filtered against a lexicon built from the training data, dropping anything the model never saw, on the reasoning that the classifier cannot be trusted on unfamiliar items.
• I3. The scope rule is picked by the cue's part of speech, so modals take their verb's subtree, raising verbs and passives add the subject, and conjunctions mark their own subtree. A single small dispatch table captures a lot of syntax.
• I4. Paths through the dependency graph between a word and the cue, plus shorter sub-paths, serve as features, letting a classifier learn graph regularities that the rules express by hand.
• I5. The authors argue that scope belongs over semantic propositions rather than continuous surface spans, and show sentences where keeping the span continuous drags certain claims into a hedged region.

For Mova

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

• I1 (learn): Self-induced repair rules could feed their verdicts to a trained corrector as features, which keeps every decision traceable to a rule plus a learned adjustment.
• I2 (frames): Speech-act or modality markers found by a tagger could be checked against a curated list of hedge or attitude words before they change a frame's certainty.
• I3 (frames): Mova could mark which propositions in the UD tree are hedged by applying rules keyed on the cue's UPOS, then attach an uncertainty flag to the frames inside that subtree.
• I4 (tree): Bounded dependency-path features could help any token-level decision that hangs on the parse, such as attaching modifiers or negation scope.
• I5 (world): The world of a text should store certainty per proposition or state, not per text span, so a hedged clause does not weaken a certain one nested inside it.