Elamur Research⌕
☰
home / papers / 1993 / One Sense per Collocation

One Sense per Collocation

Summary

Yarowsky asks whether the surrounding collocation largely determines which sense an ambiguous word takes. This parallels his earlier finding that a word tends to keep one sense within a discourse. He tests the idea on binary ambiguities with five kinds of "sense":

  • hand-tagged homographs
  • French translations taken from the Canadian Hansards
  • homophones
  • OCR-confusable word pairs
  • artificial pseudo-words

The data come from a 380 million word collection of newswire, abstracts, encyclopedias, books and smaller corpora. Collocation types include adjacent words, nearby words of a given part of speech, and syntactic pairs such as verb/object.

For each type he measures the cross-validated entropy of the sense distribution given the collocation, and how well that type alone classifies new examples. A word shows a single sense per collocation in roughly 90–99% of cases. A decision-list classifier that relies on the strongest available local cue reaches 92% precision.

The study shows that very local, position-specific evidence is more reliable than unordered wide-context "bag of words" models, especially for verbs and adjectives. This has consequences for building and checking sense-tagged data.

Key points

  • Data: a 380M-word corpus drawing on AP Newswire, Wall Street Journal, NSF/DOE abstracts, the Canadian Hansards, the Brown Corpus, ATIS/TIMIT, more than 100 books published by Harper & Row, and two encyclopedias (Grolier's and a medical one). Training and test samples were drawn from separate blocks of the corpus.
  • Entropy: cross-validated estimates correct for unseen counter-examples. A 1/0 count implies the minor sense still appears about 6% of the time. Overall entropy for aid/aide was only 0.09 bits.
  • Algorithm: a decision list (Rivest 1987) ranks evidence by the absolute log-likelihood ratio and answers from the single best matching cue, never combining probabilities.
  • Reliability of sparse evidence:
    • One unopposed word-to-the-right observation gives over 92% accuracy against a 69% prior, and accuracy is near perfect by 15/0.
    • The same kind of evidence 30 words away adds only about 3% over chance.
    • Adjacent function words help modestly, up to 13% above chance for verbs.
  • Applications:
    • Frequent collocations (above 10–15 occurrences) whose senses are mixed should be flagged as likely tagging errors.
    • Sense tagging can be sped up by assigning all examples of a frequent collocate to one sense.

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: AI & machine learning

Topics: Idioms & multiword expressions

Institutions

🎓University of PennsylvaniaPhiladelphia · United States143 works · 56 645 citations

Cite

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

@inproceedings{yarowsky1993one,
  title = {{One Sense per Collocation}},
  author = {Yarowsky, David},
  booktitle = {Human Language Technology: Proceedings of a Workshop Held at Plainsboro, New Jersey, March 21-24, 1993},
  year = {1993},
  doi = {10.3115/1075671.1075731},
  url = {https://doi.org/10.3115/1075671.1075731}
}

References

12 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Bahl, L., P. Brown, P. de Souza, R. Mercer, "A Tree-Based Statistical Language Model for Natural Language Speech Recognition," in IEEE Transactions on Acoustics, Speech, and Signal Processing, 37, 1989. find
  2. Brown, Peter, Stephen Della Pietra, Vincent Della Pietra, and Robert Mercer, "Word Sense Disambiguation using Statistical Methods," Proceedings of the 29th Annual Meeting of the Association for Computational Linguistics, 1991, pp 264-270. find our card
  3. Gale, W., K. Church, and D. Yarowsky, "One Sense Per Discourse," Proceedings of the 4th DARPA Speech and Natural Language Workshop, 1992. find
  4. Gale, W., K. Church, and D. Yarowsky, "On Evaluation of Word-Sense Disambiguation Systems," in Proceedings, 30th Annual Meeting of the Association for Computational Linguistics, 1992b. find
  5. Gale, W., K. Church, and D. Yarowsky, "A Method for Disambiguating Word Senses in a Large Corpus," in Computers and the Humanities, 1993. find
  6. Hearst, Marti, "Noun Homograph Disambiguation Using Local Context in Large Text Corpora," in Using Corpora, University of Waterloo, Waterloo, Ontario, 1991. find
  7. Leacock, Claudia, Geoffrey Towell and Ellen Voorhees "Corpus-Based Statistical Sense Resolution," in Proceedings, ARPA Human Language Technology Workshop, 1993. find
  8. Kelly, Edward, and Phillip Stone, Computer Recognition of English Word Senses, North-Holland, Amsterdam, 1975. find
  9. Resnik, Philip, "A Class-based Approach to Lexical Discovery," in Proceedings of 3Oth Annual Meeting of the Association for Computational Linguistics, 1992. find
  10. Rivest, R. L., "Learning Decision Lists," in Machine Learning, 2, 1987, pp 229-246. find
  11. Sproat, R., J. Hirschberg and D. Yarowsky "A Corpus-based Synthesizer," in Proceedings, International Conference on Spoken Language Processing, Banff, Alberta. October 1992. find
  12. Yarowsky, David "Word-Sense Disambiguation Using Statistical Models of Roget's Categories Trained on Large Corpora," in Proceedings, COLING-92, Nantes, France, 1992. find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yarowsky-1993-one-sense-per-collocation/