Elamur Research⌕
☰
home / papers / 1988 / Extraction of Semantic Information from an Ordinary English Dictionary and its Evaluation

Extraction of Semantic Information from an Ordinary English Dictionary and its Evaluation

Paper · 1988 · Jun-ichi Nakamura and Makoto Nagao

Summary

Nakamura and Nagao ask whether the definitions in a standard learner's dictionary can be mined automatically for word-to-word semantic relations, as a step toward a large lexical knowledge base. They load the magnetic-tape LDOCE (1978 edition) into a conventional relational database on a mainframe. Definitions and examples are split into separate clauses. Simple pattern-matching programs then read each definition. For verbs, they take the first verb after "to" as the Key Verb. For nouns, they take the head noun as the Key Noun, and when the definition has the form "X of Y", they treat X as a Function Noun that names the relation (is-a, part-of, membership, act, state, and so on). Following these links yields an approximate thesaurus for verbs and for nouns. To estimate its quality, they compare the noun hierarchy with LDOCE's hidden semantic markers, stored in the BOX code fields. Agreement is high for some function nouns: "act" and "state" almost always mark abstract nouns. This suggests the dictionary's coarse "abstract" class could be split further. The authors also document failures: polysemy, disjunctive definitions, compound heads and gerunds.

Key points

  • LDOCE/RDB has three tables:
    • grammar/box codes: 74,130 records
    • definitions: 84,094 records
    • example sentences: 46,122 records
  • Verb results:
    • "make" is the most frequent Key Verb (1311 uses), followed by "be" (875).
    • The verb network is mostly tree-shaped.
    • It contains cyclic definitions (e.g., welcome ↔ greet; six typical cycles are listed), which the authors liken to semantic primitives.
  • Noun results:
    • "person" is the top Key Noun (2174).
    • "type" is the top non-empty Function Noun (1064).
    • The 41 Function Nouns were chosen by hand and grouped into relation types such as is-a, part-of, membership, action, state, amount, degree and form.
  • Agreement with LDOCE semantic markers:
    • Nouns with function noun "act": 584 (97%) are marked abstract.
    • Nouns with function noun "state": 391 of 398 (98%) are marked abstract.
    • About 40% of all nouns carry only the undifferentiated "abstract" marker.
  • Observed errors:
    • The "case" sense is ambiguous (attache case vs. test case).
    • Disjunctive heads ("animal or plant") lose one branch.
    • Heads like "substance" with a compound modifier are not subclassified.
    • LDOCE's own markers are inconsistent (cloth nouns are marked Solid in some entries and Movable in others).

Interesting ideas

  • I1. The function word in a "X of Y" definition names the relation, while Y supplies the hypernym, so a short hand-built list of function nouns turns flat glosses into typed links rather than plain is-a links.
  • I2. The authors check an automatically extracted resource against an independent, hidden annotation layer that ships with the same source (the BOX code semantic markers), which gives a cheap built-in evaluation.
  • I3. Cycles in the definition graph are treated as a signal, pointing to primitive concepts that the dictionary cannot define further, rather than simply as noise.
  • I4. A function noun can be more reliable than the dictionary's coarse markers for fine distinctions, such as separating act-nouns from state-nouns inside "abstract".
  • I5. The error analysis lists specific failure types (polysemous heads, coordinated heads, compound heads, gerund heads), which work as a checklist for any gloss-based extractor.

For Mova

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

  • I1 (world): Mova could parse WordNet/Wiktionary glosses with this head-plus-function-noun rule to add part-of, member-of and state-of edges to its ontology, each edge traceable to its gloss.
  • I2 (learn): compiled knowledge (e.g., hypernyms from glosses) could be scored against an existing independent label set such as WordNet lexnames or VerbNet selectional restrictions, reporting agreement for each relation type.
  • I3 (world): detecting definitional loops in the verb hierarchy could mark candidate primitive verbs (get, go, let, stop), which frames could then use as fixed anchor predicates.
  • I4 (frames): the act/state split taken from glosses could help decide whether a nominalization introduces an event frame or a state in the text world.
  • I5 (learn): these four cases could be used as targeted negative-control tests when evaluating a gloss-to-hypernym extractor.

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

🎓Kyoto UniversityKyoto · Japan40 works · 8,235 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.English · 2,263 worksEvaluation & benchmarks · 1,220 works

Cite

Cite as Jun-ichi Nakamura and Makoto Nagao. 1988. Extraction of Semantic Information from an Ordinary English Dictionary and its Evaluation. Coling Budapest 1988 Volume 2: International Conference on Computational Linguistics.

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

@inproceedings{nakamura1988extraction,
  title = {{Extraction of Semantic Information from an Ordinary English Dictionary and its Evaluation}},
  author = {Jun-ichi Nakamura and Makoto Nagao},
  booktitle = {Coling Budapest 1988 Volume 2: International Conference on Computational Linguistics},
  year = {1988},
  doi = {10.3115/991719.991734},
  url = {https://doi.org/10.3115/991719.991734}
}

References

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

  1. ALSHAWI, H., Processing Dictionary Definitions with Phrasal Pattern Hierarchies, Computational gin-guisiic~, Vol. 13 (1987).
  2. AMSLSlt, R. A., How Do I Turn This Book On?, P~c. o] Third Annual Con]. o] the UW Centre for the NOEl), pp. 75-88 (1987). find
  3. BOGURAEV, B., Experiences with a Machine-Readable Dictionary, P~vc. of Third Annual Con]. o] the UW Centre ]or the NOEl), pp. 37-50 (1987). find
  4. NAGAO, M., TsuJII, J., UEDA, Y., TAKIYAMA, M., An Attempt to Computerized Dictionary Data Bases, Proc. of COLING80, pp. 534-542 (1980). find
  5. NAKAMURA, J., FUJIGAKI, M., NAGAO, M., Longman Dictionary Database and Extraction of its Information, Report on Cognitive Approaches for Discourse Modeling , Kyoto University (1986) (in Japanese). find
  6. NAKAMURA, J., SAKAI, K., NAGAO, M., Automatic Analysis of Semantical Relation between English Nouns by an Ordinary English Dictionary, Im stitute of Electronics, Information and Communication Engineers of Japan, WGNLC, 86-23 (1987) (in Japanese). find
  7. MICHI~LS, A., NOi~L, J., Approaches to Thesaurus Production, Proc. o] COLING82, pp. 227-232 (1982). find
  8. PltOCTI~R, P., Longman Dictionary of Contemporary English Longman Group Lhnited, Harlow and London, England (1978). find
  9. .ScHANK, R. C., Conceptual Information Processing, New York, North Holland (1975). find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nakamura-1988-extraction-semantic-information-ordinary-english/