Elamur Research⌕
☰
home / papers / 2008 / Hedge Classification in Biomedical Texts with a Weakly Supervised Selection of Keywords

Hedge Classification in Biomedical Texts with a Weakly Supervised Selection of Keywords

Summary

The paper tackles hedge detection, which means deciding whether a sentence in a biomedical text states a fact or expresses speculation. Speculative statements tend to show up as false positives in downstream tasks such as gene interaction extraction and ICD-9-CM coding of radiology reports. Szarvas builds on the weakly supervised dataset of Medlock and Briscoe (2007), whose training sentences were collected by iterative bootstrapping from FlyBase articles.

Bootstrapping skews word statistics in a harmful way: irrelevant words that often co-occur with true hedge cues also start to look speculative. To counter this, the author designs a multi-step keyword selection procedure. It ranks candidate uni-, bi- and trigrams, credits only the strongest cues in each sentence, and removes longer phrases that a shorter substring already covers. A brief manual check then finalizes the feature set for a Maximum Entropy classifier.

In the parts of the paper available here, the main finding is that hedge detection reduces largely to finding task-specific keywords. Classifiers built this way port poorly to texts from other sources, such as BMC Bioinformatics, because the same cue words are used differently there.

Key points

  • In the annotated test set, 638 of 1968 gene name mentions fall in speculative sentences. That is roughly one gene in three that should be excluded from interaction extraction.
  • In an ICD-9-CM coding system, adding a simple hedge lookup module with 38 keywords raised overall performance from 79.7% to 89.3%.
  • Bootstrapping distorts statistics for irrelevant words. For example, P(spec|it) rose from 74.67% on the seed sets to 94.32% after the training data was extended.
  • The selection procedure reduced 2407 candidates to 362 and then to 253. This removed 89.5% automatically and left only about 5 minutes of expert validation.
  • On BMC Bioinformatics texts, just two keywords (possible, likely) caused half of the false positives, which limits cross-source portability. The author suggests bi- and trigrams that capture non-speculative uses as a remedy.

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: Linguistics & languages

Cited by

30 papers of our library cite it (of the 140673 whose bibliographies we read); 5 of them have a page here, 25 not yet.

Cite

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

@inproceedings{szarvas2008hedge,
  title = {{Hedge Classification in Biomedical Texts with a Weakly Supervised Selection of Keywords}},
  author = {Szarvas, György},
  booktitle = {Proceedings of ACL-08: HLT},
  year = {2008},
  url = {https://aclanthology.org/P08-1033/}
}

References

11 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Adam L. Berger, Stephen Della Pietra, and Vincent J. Della Pietra. 1996. A maximum entropy approach to natural language processing. Computational Linguistics, 22(1):39–71. ↗ our card
  2. Wendy W. Chapman, Will Bridewell, Paul Hanbury, Gregory F. Cooper, and Bruce G. Buchanan. 2001. A simple algorithm for identifying negated findings and diseases in discharge summaries. Journal of Biomedical Informatics, 5:301–310. ↗
  3. Ken Hyland. 1994. Hedging in academic writing and eap textbooks. English for Specific Purposes, 13(3):239– 256. find
  4. Marc Light, Xin Ying Qiu, and Padmini Srinivasan. 2004. The language of bioscience: Facts, speculations, and statements in between. In Lynette Hirschman and James Pustejovsky, editors, HLT-NAACL 2004 Workshop: BioLINK 2004, Linking Biological Literature, Ontologies and Databases, pages 17–24, Boston, Massachusetts, USA, May 6. Association for Computational Linguistics. find
  5. Ben Medlock and Ted Briscoe. 2007. Weakly supervised learning for hedge classification in scientific literature. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 992–999, Prague, Czech Republic, June. Association for Computational Linguistics. find our card
  6. Yusuke Miyao and Jun’ichi Tsujii. 2005. Probabilistic disambiguation models for wide-coverage HPSG parsing. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL’05), pages 83–90, Ann Arbor, Michigan, June. Association for Computational Linguistics. find our card
  7. Marie A. Moisio. 2006. A Guide to Health Insurance Billing. Thomson Delmar Learning. find
  8. John P. Pestian, Chris Brew, Pawel Matykiewicz, DJ Hovermale, Neil Johnson, K. Bretonnel Cohen, and Wlodzislaw Duch. 2007. A shared task involving multi-label classification of clinical free text. In Biological, translational, and clinical language processing, pages 97–104, Prague, Czech Republic, June. Association for Computational Linguistics. find
  9. Ellen Riloff, Janyce Wiebe, and Theresa Wilson. 2003. Learning subjective nouns using extraction pattern bootstrapping. In Proceedings of the Seventh Computational Natural Language Learning Conference, pages 25–32, Edmonton, Canada, May-June. Association for Computational Linguistics. find our card
  10. James G. Shanahan, Yan Qu, and Janyce Wiebe. 2005. Computing Attitude and Affect in Text: Theory and Applications (The Information Retrieval Series). Springer-Verlag New York, Inc., Secaucus, NJ, USA. find
  11. Janyce Wiebe, Theresa Wilson, Rebecca F. Bruce, Matthew Bell, and Melanie Martin. 2004. Learning subjective language. Computational Linguistics, 30(3):277–308. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/szarvas-2008-hedge-classification-biomedical-texts-weakly/