📄 Automatic Identification of Non-compositional Phrases

Summary

The paper tackles the automatic detection of non-compositional phrases (idioms such as "red tape" or "spill gut"). Their meaning cannot be built from their parts, which can mislead machine translation, query expansion in information retrieval, and similarity-based word sense disambiguation. Lin's core idea is that an idiomatic phrase behaves differently in distribution from phrases made by swapping one of its words for a near-synonym.

A large newspaper corpus was parsed with Minipar. Dependency triples were filtered into a collocation database, and a corpus-based thesaurus supplied similar words for each head and modifier. For every collocation, mutual information is computed along with a confidence interval derived from a normal approximation to the binomial frequency count. A phrase is labelled non-compositional when no substituted variant exists whose 95% interval overlaps its own.

Comparison with WordNet was not feasible because its multiword entries are already single lexical items in the parser. The output was therefore judged by manually inspecting results for a sample of words. The approach matters because it supplies idiom lists without hand-built resources.

Key points

• Input data: a 125-million-word newspaper corpus parsed with Minipar (a descendant of Principar). This yielded roughly 80 million dependency relationships, which were reduced to about 11 million unique collocations using a log-likelihood ratio threshold of 0.5.
• Thesaurus: built with Lin's (1998) similarity measure, it covers 11,839 nouns, 3,639 verbs and 5,658 adjectives/adverbs that occur at least 100 times.
• Illustrative cases:
  • "spill gut" (13 occurrences, MI 6.24) has no attested substitutes.
  • "red tape" (259 occurrences, MI 5.87) far exceeds "yellow tape" (MI 3.75).
  • Compositional "economic impact" has many variants with comparable MI values.
• Significance test: confidence intervals use z = 1.96 for the 95% level. For example, "make difference" (MI 2.928) and "make change" (MI 2.194) have disjoint intervals, so their MI values count as significantly different.
• Evaluation sample:
  • The five most frequent open-class words: have, company, make, do, take.
  • Five words with frequency ranks 2000–2004, such as path, lock, resort and column.
  • An appendix also lists results obtained without a parser, using word proximity.