Extended Gloss Overlaps as a Measure of Semantic Relatedness

Summary

The authors propose a measure of semantic relatedness between concepts that relies on dictionary definitions (glosses). The method builds on the Lesk algorithm, which counts the words shared by definitions. However, the glosses in WordNet are too short to distinguish fine degrees of relatedness. The extended gloss overlaps measure therefore compares not only the glosses of the two synsets but also the glosses of concepts linked to them by explicit WordNet relations: hypernyms, hyponyms, meronyms, holonyms, attribute, similar-to, also-see. Usage examples are also included in the comparison. The method looks for the longest shared word sequences, and an overlap of n words scores n² points, since longer phrases are less likely to match. The measure was evaluated in two ways. First, it was compared with human judgments of relatedness. Then it was used to build a word sense disambiguation (WSD) algorithm that selects the sense of a target word based on its relatedness to the neighboring words in the context. On the SENSEVAL-2 lexical sample data, this algorithm was outperformed by only one of the participating systems. Importantly, the measure works not only for nouns but also for verbs and adjectives. Moreover, the work shows that the richer the network of relations in a dictionary, the more accurately relatedness can be estimated.

Key points

The correlation of the measure with human relatedness judgments is at least 0.6.
The average gloss length in WordNet is only about seven words. The relations used cover more than 93% of all links in WordNet.
Overlap scoring: a phrase of n words scores n² points. Function words (pronouns, prepositions, articles, conjunctions) cannot appear at the beginning or end of an overlap.
In WSD on SENSEVAL-2 with all relation pairs, the F-measure is 0.42 for nouns, 0.35 for verbs and 0.26 for adjectives. No single minimal set of relation pairs comes close to these values.
The most useful relations: for nouns, hyponyms and meronyms; for adjectives, also-see and attribute; for verbs, hyponyms. Usage examples help for all parts of speech. In a Lesk-like baseline on SENSEVAL-2, they raised recall from 16% to 23%.

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

Institutions

🎓Carnegie Mellon University

Pittsburgh · United States278 works · 138 810 citations 🎓University of Minnesota, Duluth

Duluth · United States2 works · 1 148 citations

Cite

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

@inproceedings{banerjee2003extended,
  title = {{Extended Gloss Overlaps as a Measure of Semantic Relatedness}},
  author = {Banerjee, Satanjeev and Pedersen, Ted},
  booktitle = {IJCAI-03},
  year = {2003},
  url = {https://www.ijcai.org/Proceedings/03/}
}

References

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

A. Budanitsky and G. Hirst. Semantic distance in WordNet: An experimental, application-oriented evaluation of five measures. In Workshop on WordNet and Other Lexical Resources, Second meeting of the North American Chapter of the Association for Computational Linguistics, Pittsburgh, June 2001. find
Y. Choueka and S. Lusignan. Disambiguation by short contexts. Computers and the Humanities, 19:147-157, 1985. lEdmonds and Cotton, 2001] P. Edmonds and S. Cotton, editors. Proceedings of the Senseval-2 Workshop. Association for Computational Linguistics, Toulouse, France, 2001. find
C. Fellbaum, editor. WordNet: An electronic lexical database. M I T Press, 1998. Uiang and Conrath, 1997] J. Jiang and D. Conrath. Semantic similarity based on corpus statistics and lexical taxonomy. In Proceedings on International Conference on Research in Computational Linguistics, Taiwan, 1997. L… find
D. Lin. Using syntactic dependency as a local context to resolve word sense ambiguity. In Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics, pages 64-71, Madrid, July 1997. find
G.A. Miller and W.G. Charles. Contextual correlates of semantic similarity. Language and Cognitive Processes, 6(1): 1-28, 1991. find
P. Resnik. Using information content to evaluate semantic similarity in a taxonomy. In Proceedings of the 14th International Joint Conference on Artificial Intelligence, Montreal, August 1995. find our card
H. Rubenstein and J.B. Goodenough. Contextual correlates of synonymy. Computational Linguistics, 8:627-633, 1965. find
G. Zipf. The Psycho-Biology of Language. Houghton Mifflin, Boston, M A , 1935. find

This page in Markdown → https://md.research.elamur.ai/free/en/papers/banerjee-2003-extended-gloss-overlaps-as-measure/

📄 Extended Gloss Overlaps as a Measure of Semantic Relatedness