📄 Features for Detecting Hedge Cues

Summary

The paper addresses the CoNLL-2010 shared task 1, biological track: deciding whether a biomedical sentence states a fact or something uncertain by finding hedge cues such as "may", "putative" or "suggest". Cue detection is framed as BIO sequence tagging (B-CUE, I-CUE, O). A sentence counts as uncertain if at least one cue is tagged in it.

The tagger is a discriminative Markov model decoded with Viterbi. Its weights are learned with Bayes Point Machines (BPM), which average several perceptrons, each trained on a different shuffle of the data. Features include:

• surface properties of each token;
• GENIA tagger output (POS, chunks, named entities);
• Porter stems;
• a discretized score of how distributionally similar a stem is to known hedge cues, computed on the unlabeled corpus.

A genetic algorithm searched for useful feature combinations. BPM reached a sentence-level F1 of 85.36, slightly above CRF and plain Perceptron, and it trained faster than CRF. That speed made the repeated feature search affordable.

Key points

• Data and tools: the CoNLL-2010 biological labeled and unlabeled corpora, the GENIA tagger (Penn Treebank-style tokens, POS, chunk and NE tags), the Porter stemmer, and the authors' C++ structured-prediction library (soplib).

• Distributional feature: for each stem, a histogram of neighbouring words (window of 3) is built from the unlabeled corpus. It is compared with the histograms of training hedge cues using the Tanimoto coefficient. The highest similarity is then turned into binary threshold flags from >0.9 down to >0.1 (F16–F24).

• Learner comparison on sentence-level F1:

  Learner	P	R	F1
  BPM (T=5, K=20)	89.12	81.90	85.36
  CRF (unpenalized beat C=1)	—	—	85.11
  Perceptron (T=40)	83.98	84.94	84.46

  Perceptron trades lower precision for higher recall.

• With T fixed at 5, BPM performance levels off once about 20 perceptrons are averaged; adding more brings no further gain.

• Ablation of each feature-table row on the test set:

  • Two large window-combination templates turned out to overfit, and dropping one of them improved results.
  • Every other template helped, including non-intuitive combinations involving suffixes, POS, chunks and similarity flags.
  • The last-letter (F9) feature template affected predictions for cues such as "believe", "putative" and "presumably".

Interesting ideas

• I1. Lexicon expansion from seeds: a word's context-histogram similarity to a small set of known cue words, computed on raw text, can be turned into cumulative threshold flags, giving a soft and corpus-driven lexicon feature.
• I2. Derive the sentence label from spans: the system tags cue spans and calls a sentence uncertain whenever any cue appears, so every sentence-level decision points back to the concrete words that caused it.
• I3. Genetic search over feature conjunctions on a three-way split: one part trains, one scores fitness, and a third, rarely consulted part guards against the search overfitting its own fitness set.
• I4. Per-template ablation as an overfitting check: removing each feature group in turn revealed which groups helped and which only memorized the training data.
• I5. Cheap ensemble by averaging shuffled perceptrons: averaging perceptrons trained on different sample orders matched CRF quality at lower cost, and gains stopped at around 20 members.

For Mova

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

• I1 (tokens): this could flag candidate epistemic or hedging words in the English lexicon from unlabeled text, with each flag traceable to its nearest seed cue.
• I2 (frames): speech-act and certainty marking in prag could attach an explicit hedge-cue node to a frame, so "may inhibit" becomes a hypothetical claim with an inspectable reason.
• I3 (learn): self-induced repair rules or feature templates could be evolved this way, with the third split acting as an untouched control.
• I4 (learn): routine leave-one-template-out runs would expose parser or tagger features that only fit silver data, complementing negative controls.
• I5 (tree): the perceptron parser could average a small fixed number of shuffled runs for steadier weights while keeping the transparent linear model.