Elamur Research⌕
☰
home / papers / 2000 / Integrating Subject Field Codes into WordNet

Integrating Subject Field Codes into WordNet

Summary

WordNet lacks systematic domain information: dictionary-style field labels such as MEDICINE appear only sporadically and inconsistently. Yet NLP tasks like IR, WSD and text classification benefit from knowing a word sense's subject area. The authors built a resource that labels WordNet 1.6 synsets with Subject Field Codes (SFCs).

They first collected a code inventory from dictionaries and the Dewey Decimal Classification. They then arranged it into a hierarchy and into flexible "families" of related fields. Annotation was semi-automatic and iterative:

  • a few high-level synsets were labelled by hand;
  • the labels were spread along WordNet relations, with explicit exceptions that block wrong inheritance;
  • errors found in a text classification test were used to fix the manual seeds.

By the time of writing, about 96% of noun synsets were covered by 115 codes. When the codes were used to classify Adnkronos news, the test set reached 0.95 precision and 0.96 recall. The resource groups senses across WordNet sub-hierarchies and parts of speech. The authors argue it is largely language independent, which makes it useful for multilingual wordnets.

Key points

  • Code inventory:
    • roughly 250 initial SFCs, taken from paper and machine-readable dictionaries and extended using the Dewey Decimal Classification;
    • 115 codes actually used, in a four-level hierarchy;
    • 41 second-level codes treated as the most informative.
  • Propagation relations: hyponymy, troponymy, meronymy, antonymy and pertain-to. Exceptions prevent unwanted spread; for example, parts of a barbershop should not inherit COMMERCE.
  • Coverage:
    • about 63,000 of 66,000 noun synsets annotated, with 72,763 synset/SFC mappings;
    • 95% of synsets have one code, 3% have two, and 2% have three (mostly due to regular polysemy).
  • FACTOTUM code: collects domain-neutral synsets, comprising 2,780 "stop senses" (numbers, weekdays, colours) and 3,670 generic senses, 6,450 synsets in total.
  • Evaluation: text classification on English translations of Italian Adnkronos news, using nouns, proper nouns and abbreviations only.
    • Training: 0.92 precision, 0.96 recall.
    • Test: 0.95 precision, 0.96 recall.
    • Weak codes such as PSYCHOLOGY had low per-code precision (0.70).

Interesting ideas

  • I1. Seed-and-propagate labelling with explicit exceptions. A small set of hand labels is spread through a lexical graph, and declarative blocking rules stop it where it would go wrong. This would scale domain tagging of a Ukrainian wordnet or thesaurus with little manual work.
  • I2. Effort metrics per category. "Productivity" compares the synsets reached after inheritance with the number of manual labels. "Relevance" is a category's share of all annotations. Together they show which domains are scattered across the hierarchy and therefore costly, which helps when planning annotation budgets.
  • I3. Extrinsic task as a diagnostic tool. The classification test was used mainly to locate annotation errors, not to compete as a classifier:
    • repeated misclassification into one code suggests over-propagation or a misplaced code;
    • a code that is never predicted suggests missing low-level assignments.
  • I4. A catch-all FACTOTUM label. Domain-neutral and stop-like senses get their own label instead of being forced into some field. This keeps domain signals clean, a design choice applicable to any topic-labelled lexicon.
  • I5. Fixed hierarchy plus flexible families. The inclusion hierarchy stays stable, while application-specific groupings (e.g., SPORT + MEDICINE + ANATOMY) can be recombined freely. This separates a stable schema from user-specific views.

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

Topics: Word sense disambiguation

Cited by

101 papers of our library cite it (of the 140673 whose bibliographies we read); 6 of them have a page here, 95 not yet.

Cite

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

@inproceedings{magnini2000integrating,
  title = {{Integrating Subject Field Codes into WordNet}},
  author = {Magnini, Bernardo and Cavaglià, Gabriela},
  booktitle = {Proceedings of the Second International Conference on Language Resources and Evaluation (LREC'00)},
  year = {2000},
  doi = {10.63317/36otuawjvvk5},
  url = {https://doi.org/10.63317/36otuawjvvk5}
}

References

17 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Agirre, E., G. Rigau, L. Padro, and J. Atserias, to appear. Combining supervised and unsupervised lexical knowledge methods for word sense disambiguation. Computers and the Humanities. Special Issue on SensEval. To appear. find
  2. Artale, Alessandro, Anna Goy, Bernardo Magnini, Emanuele Pianta, and Carlo Strapparava, 1998. Coping with wordnet sense proliferation. In First International Conference on Language Resources & Evaluation. Granada, Spain. find
  3. Diekema, Anne, 1998. Dewey decimal classification. Http://istweb.syr.edu/ isdp561/Dewey/dui.html. find
  4. Dyvik, Helge, 1998. Translation as semantic mirrors. In The 13th Biennal European Conference on Artificial intelligence. ECAI’98. find
  5. Gonzalo, Julio, Felisa Verdejo, Carlos Peters, and Nicoletta Calzolari, 1998. Applying eurowordnet to cross-language text retrieval. Computers and the Humanities, 32(2-3):185–207. find
  6. Grefenstette, Gregory, 1994. Explorations in Automatic Thesaurus Discovery. Kluwer Academic Publishers. find
  7. Gythrie, Joe A., Louise Guthrie, Yorick Wilks, and Homa Aidinejad, 1991. Subject-dependent co-occurrence and word sense disambiguation. In 29th Annual meeting of the Association for Computational Linguistics. Berkeley, California, USA. find
  8. Hearst, Marti A. and Hinrich Schutze, 1996. Corpus Processing for Lexical Acquisition, chapter Customizing a Lexicon to Better Suit a Computational Task. The MIT Press, pages 77–96. find
  9. Leacock, Claudia, Geoffrey Towell, and Ellen M. Voorhees, 1996. Corpus Processing for Lexical Acquisition, chapter Towards Building Contextual representaions of WordSenses Using Statistical Models. The MIT Press, pages 77–96. find
  10. Liddy, Elizabeth and Woojin Paik, 1993. Document filtering using semantic information from a machine readable dictionary. In Workshop on Very Large Corpora. ACL. find
  11. Lin, Dekang, 1998. Automatic retrieval and clusterind of similar words. volume II. Montreal: COLING-ACL’98. find
  12. Miller, George A., Richard Beckwith, Christiane Fellbaum, Derek Gross, and Katherine J. Miller, 1990. Wordnet: An on-line lexical database. International Journal of Lexicography, 3(4):235–312. find
  13. Pustejovsky, James, 1995. The Generative Lexicon. Massachusetts Institute of Technology. find
  14. Schutze, Hinrich, 1992. Dimensions of meaning. minneapolis, MN: Supercomputing’92. find
  15. Schutze, Hinrich, 1998. Automatic word sense discrimination. Computational Linguistics, 24(1):97–123. find our card
  16. Veronis, Jean and Nancy Ide, 1990. Word sense disambiguation with very large neural networks extracted from machine readable dictionaries. In Proceedings of the 13th international Conference on Computational Linguistics, volume 2. helsinki: COLING’90. find
  17. Walker, D. E. and R. A. Amsler, 1986. Analyzing Language in Restricted Domain. Sublanguage description and Processing, chapter The use of Machine Readable Dictio- find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/magnini-2000-integrating-subject-field-codes-into/