Elamur Research⌕
☰
home / papers / 2010 / Exploiting Multi-Features to Detect Hedges and their Scope in Biomedical Texts

Exploiting Multi-Features to Detect Hedges and their Scope in Biomedical Texts

Paper · 2010 · Huiwei Zhou, Xiaoyan Li, Degen Huang et al.

Summary

The paper addresses speculation in biomedical writing. It covers two CoNLL-2010 Shared Task problems: flagging sentences that contain uncertain claims (Task 1) and marking which words each hedge cue governs (Task 2). The authors treat sentence detection as cue detection: a sentence counts as uncertain if it has at least one cue.

Cue detection combines two steps. First, a lexicon is matched against the text using longest-match search. The lexicon holds every cue seen in training, optionally expanded with WordNet synonyms. Second, a CRF tagger with IOB2 labels uses the lexicon matches as features, alongside words, stems, POS and chunk labels from GENIA Tagger. For scope, a CRF that labels the first and last scope tokens is compared with a system built on syntactic patterns.

Cue detection reaches 86.32% sentence-level F-score in-domain. Scope finding is much weaker: 44.42% officially, rising to 54.18% after a post-hoc fix rule and extra cue-distance features. The work shows that a cheap lexicon plus a sequence model handles cue detection well, while scope resolution remains the hard part.

Key points

  • The training data yielded 143 distinct hedge cues: 90 unigrams, 24 bigrams, and longer n-grams up to five tokens. Adding WordNet synonyms grows the keyword list to 438 entries, 359 of them single words.
  • Maximum matching picks the longest cue available. For example, "indicate that" wins over plain "indicate".
  • Task 1 results: the submitted system scored 85.77 F. The best configuration, using synonyms and tuned on context window size, reached 86.32.
  • Task 2 official in-domain results: the CRF scored 44.42 F and the syntactic-pattern system 43.45. A rule that moves the scope end so it includes trailing parenthesised material raised the CRF to 51.83.
  • Further Task 2 gains came from features describing the nearest preceding and following cue: its distance, stem and POS. With a window of 4 these gave 54.18. With gold cues the CRF reaches 59.48, which serves as an upper bound.

Interesting ideas

  • I1. Lexicon output used as a feature, not a verdict: dictionary hits become IOB labels inside the CRF, so the model learns when an ambiguous cue word is not actually hedging
  • I2. Longest-match lookup over a multiword cue inventory ensures that overlapping entries such as "possibility", "raise the possibility" and "cannot rule out the possibility" resolve to the fullest expression
  • I3. When the relevant anchor can fall outside a fixed context window, the authors add explicit features for distance to the nearest anchor and that anchor's stem and POS, which recovers long-range information cheaply
  • I4. Error analysis showed that predicted scopes systematically ended just before a closing parenthesis, and a single repair rule fixed a large share of the errors
  • I5. The best upstream cue detector did not give the best downstream scope result, and a gold-cue run separated cue errors from scope errors

For Mova

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

  • I1 (frames): prag could feed a hand-built list of modality and hedge cues into its speech-act and implicature classifier as evidence, letting context decide whether "suggest" or "may" really marks uncertainty.
  • I2 (tokens): the en crate could apply greedy longest-match to multiword expressions and fixed phrases before tagging, so that later stages see one unit instead of fragments.
  • I3 (frames): scope of hedges or negation over a frame could be computed from tree distance to the nearest cue node, so that "presumed" can mark a whole embedded clause as non-factual in the world model.
  • I4 (learn): this is a good template for self-induced repair rules, where recurring boundary mismatches against gold or silver data are mined and turned into explainable corrections.
  • I5 (learn): pipeline evaluation should report each stage both with predicted and with gold upstream input, and should not assume that improving one stage improves the final answer.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Arts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Huiwei Zhou, Xiaoyan Li, Degen Huang, Zezhong Li and Yuansheng Yang. 2010. Exploiting Multi-Features to Detect Hedges and their Scope in Biomedical Texts. Proceedings of the Fourteenth Conference on Computational Natural Language Learning – Shared Task.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{zhou2010exploiting,
  title = {{Exploiting Multi-Features to Detect Hedges and their Scope in Biomedical Texts}},
  author = {Huiwei Zhou, Xiaoyan Li, Degen Huang, Zezhong Li and Yuansheng Yang},
  booktitle = {Proceedings of the Fourteenth Conference on Computational Natural Language Learning – Shared Task},
  year = {2010},
  url = {https://aclanthology.org/W10-3015/}
}

References

14 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Leonard E. Baum, and Ted Petrie. 1966. Statistical inference for probabilistic functions of finite state Markov chains. Annals of Mathematical Statistics , 37(6):1554–1563. ↗
  2. Peter L. Elkin, Steven H. Brown, Brent A. Bauer, Casey S. Husser, William Carruth, Larry R. Bergstrom, and Dietlind L. Wahner-Roedler. 2005. A controlled trial of automated classification of negation from clinical notes. BMC Medical Informatics and Decision Making , 5(13). ↗
  3. Richárd Farkas, Veronika Vincze, György Móra, János Csirik, and György Szarvas. 2010. The CoNLL-2010 Shared Task: Learning to Detect Hedges and their Scope in Natural Language Text. In Proceedings of CoNLL-2010: Shared Task , 2010, pages 1–12.
  4. Viola Ganter, and Michael Strube. 2009. Finding hedges by chasing weasels: Hedge detection using wikipedia tags and shallow linguistic features. In Proceedings of the ACL-IJCNLP 2009 Conference Short Papers , pages 173–176.
  5. Dan Klein, and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the 41st Meeting of the Association for Computational Linguistics , pages 423–430.
  6. John Lafferty, Andrew McCallum, and Fernando Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference Learning , pages 282–289. on Machine Marc Light, Xin Ying Qiu, and Padmini Srinivasan. 2004. The language of bioscience: facts, speculations, and statements in between. In HLT-NAACL 2004 Workshop: BioLINK 2004, Linking Biological Literature, Ontologies and Databases , pages 17–24. find
  7. Yuan Liu, Qiang Tan, and Kunxu Shen. 1994. The word segmentation rules and automatic word segmentation methods for Chinese information processing . QingHua University Press and GuangXi Science and Technology Press. find
  8. Andrew McCallum, Dayne Freitag, and Fernando Pereira. 2000. Maximum entropy Markov models for information extraction and segmentation. In Proceedings of ICML 2000 , pages 591–598 . find
  9. Ben Medlock. 2008. Exploring hedge identification in biomedical literature. Journal of Biomedical Informatics , 41(4):636–654. ↗
  10. Ben Medlock, and Ted Briscoe. 2007. Weakly supervised learning for hedge classification in scientific literature. In Proceedings of ACL-07, pages 992–999.
  11. Roser Morante, and Walter Daelemans. 2009. Learning the scope of hedge cues in biomedical texts. In Proceedings of the Workshop on BioNLP, ACL 2009 , pages 28–36.
  12. Ellen Riloff, Janyce Wiebe, and Theresa Wilson. 2003. Learning subjective nouns using extraction pattern bootstrapping. In Proceedings of the 7th Conference on Computational Language Learning, pages 25–32.
  13. Natural György Szarvas. 2008. Hedge classification in biomedical texts with a weakly supervised selection of keywords. In Proceedings of ACL: HLT , pages 281–289.
  14. György Szarvas, Veronika Vincze, Richárd Farkas, and János Csirik. 2008. The BioScope corpus: biomedical texts annotated for uncertainty, negation and their scopes. In Proceedings of BioNLP 2008: Current Trends in Biomedical Language , pages 38–45. Natural Yoshimasa Tsuruoka, Yuka Tateishi, Jin-Dong Kim, Tomoko Ohta, John McNaught, Sophia Ananiadou, Jun´ichi Tsujii. 2005. Developing a robust part-of-speech tagger for biomedical text. In Advances in Informatics 2005 , pages 382–392. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhou-2010-exploiting-multi-features-detect-hedges/