Elamur Research⌕
☰
home / papers / 2003 / Learning Semantic Constraints for the Automatic Discovery of Part-Whole Relations

Learning Semantic Constraints for the Automatic Discovery of Part-Whole Relations

Paper · 2003 · Roxana Girju, Adriana Badulescu and Dan Moldovan

Summary

The paper tackles automatic detection of part-whole (meronymy) relations in English text. The most frequent surface patterns for this relation, such as genitives and "have"-type verbs, are highly ambiguous. The authors first gathered meronymic lexico-syntactic patterns semi-automatically. They took 100 WordNet part-whole pairs and searched for them in SemCor 1.7 and LA Times (TREC-9) sentences. They then kept the three most frequent patterns: "X of Y", "Y's X" and "Y verb X". To separate true part-whole readings from other meanings, they built a large annotated training set from parsed corpus sentences, with nouns sense-tagged in WordNet and extra positive pairs taken from WordNet itself. C4.5 decision trees were trained on the WordNet semantic classes of the part and the whole. These classes were refined step by step by moving down the IS-A hierarchy. The outcome was 20 semantic constraints, which detected part-whole relations with about 83% accuracy on a 10,000-sentence corpus. This beats earlier statistical work and is presented as a method that carries over to other semantic relations.

Key points

  • Pattern discovery found 535 meronymic occurrences in 20,000 sentences. Of these, 92.15% were phrase-level patterns. There were 54 distinct patterns: 36 phrase-level and 18 sentence-level.
  • Pattern frequencies:
    • "X of Y" covered 35% of phrase-level cases.
    • "Y's X" covered 14% of phrase-level cases.
    • "Y verb X" covered 43% of sentence-level cases.
  • Training data came from 19,000 SemCor sentences and 100,000 LA Times sentences, parsed with the Charniak parser. They yielded 53,944 pattern matches: 6,973 part-whole and 46,971 not. Adding 27,636 WordNet pairs gave 34,609 positives in total.
  • The learner was C4.5. Its 20 resulting rules restrict the WordNet classes of part and whole, including negated classes (e.g., an object that is not a substance).
  • Overall accuracy was about 83%, compared with 55% on the top 50 parts reported by Berland and Charniak. A Question Answering use case showed "answer fusion" building a parts hierarchy for the AH-64A Apache helicopter.

Interesting ideas

  • I1. Seed-pair pattern mining. Start from known related pairs in a lexical resource and search corpora to see which surface patterns link them. This bootstraps pattern inventories for any relation or language that has a wordnet-like resource.
  • I2. Generalize, then specialize. Map words to top-level ontology classes, then descend the IS-A hierarchy only where examples stay mixed. This yields compact, human-readable constraints instead of memorized word pairs.
  • I3. Positives from the knowledge base. Mixing ontology pairs with corpus examples enlarges the set of positive examples cheaply. This is useful when annotated positives are rare compared with negatives.
  • I4. Longest-match concept lookup. Keep the longest word sequence of a noun phrase that exists as a dictionary concept (e.g., "carving knife" from "brown carving knife"). This is a simple, reusable normalization step before semantic lookup.
  • I5. Answer fusion. Assemble relation instances scattered across documents into one hierarchy to answer composition-type questions. This applies to any QA or knowledge-extraction pipeline in open domains.

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

 Institutions

🎓The University of Texas at DallasRichardson · United States47 works · 5,363 citations

📚 Shelves 4 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works

Cite

Cite as Roxana Girju, Adriana Badulescu and Dan Moldovan. 2003. Learning Semantic Constraints for the Automatic Discovery of Part-Whole Relations. Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics.

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

@inproceedings{girju2003learning,
  title = {{Learning Semantic Constraints for the Automatic Discovery of Part-Whole Relations}},
  author = {Roxana Girju, Adriana Badulescu and Dan Moldovan},
  booktitle = {Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2003},
  doi = {10.3115/1073445.1073456},
  url = {https://doi.org/10.3115/1073445.1073456}
}

References

11 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Matthew Berland and Eugene Charniak. 1999. Finding Parts in Very Large Corpora. In Proceedings of the 37th Annual Meeting of the Association for the Computational Linguistics (ACL-99), pages 57-64, College Park, MD. our card
  2. Eugene Charniak. 2000. A maximum-entropy-inspired parser. In Proceedings of the North American Chapter of the Association for Computational Linguistics (NAACL 2000), Seattle, WA. our card
  3. M. Evens, B. Litowitz, J. Markowitz, R. Smith, and O. Werner. 1980. Lexical-semantic relations: A Comparative Survey. In Proceedings of the Linguistic Research, pages 187-190, Edmonton, Canada. find
  4. Peter Gerstl and Simone Pribbenow. 1996. A conceptual theory of Part-Whole Relations and Its Applications. In Data and Knowledge Engineering, 20(3), pages 305-322. find
  5. Roxana Girju. 2001. Answer Fusion with On-Line Ontology Development. In Proceedings of the North American Chapter of the Association for Computational Linguistics (NAACL) - Student Research Workshop, (NAACL 2001), Pittsburgh, PA. find
  6. Marti Hearst. 1998. Automated Discovery of Word-Net Relations. In An Electronic Lexical Database and Some of its Applications, MIT Press, Cambridge, MA. find
  7. Madelyn Iris, Bonnie Litowitz and Martha Evens. 1988. Problems of the Part-Whole Relation. In Relational Models of the Lexicon: Representing Knowledge in Semantic Networks, Cambridge University Press, Evens (ed.), pages 261-288. find
  8. Rada Mihalcea and Dan Moldovan. 2001. A Highly Accurate Bootstrapping Algorithm for Word Sense Disambiguation. In the International Journal on Artificial Intelligence Tools, vol.10, no.1-2, pages 5-21. find
  9. S. Harabagiu, D. Moldovan, M. Pasca, M. Surdeanu, R. Mihalcea, R. Girju, V. Rus, F. Lacatusu, P. Moraescu, and R. Bunescu. 2001. Answering Complex, List and Context Questions with LCC’s Question-Answering Server. In Proceedings of the TExt Retrieval Conference for Question Answering (TREC 10). find
  10. J. Ross Quinlan. 1993. C4.5: Programs for Machine Learning. Morgan Kaufmann. our card
  11. Morton Winston, Roger Chaffin, and Douglas Herrmann. 1987. A Taxonomy of Part-Whole Relations. In Cognitive Science, vol. 11, pages 417-444. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/girju-2003-learning-semantic-constraints-automatic-discovery/