Oxymoron generation using an association word corpus and a large-scale N-gram corpus

Abstract

Oxymorons are combinations of contradictory or incongruous words and are typically used to draw readers' attention to a text. This paper proposes a method to generate oxymorons using an association word corpus and a large-scale $$N$$N-gram corpus. First, adjectives are fed as input into the proposed system. Then, antonym pairs are extracted from the $$N$$N-gram corpus. Using the antonym pairs and the association word corpus, candidates are generated. Candidates are finalized or eliminated according to their suitability and attractiveness. To determine suitability, pointwise mutual information (PMI) is employed to exclude grammatically unnatural expressions. To determine attractiveness, PMI, gap of frequency of the oxymoron candidates, and WordNet are used. The generated combinations of oxymorons indicate the potential effectiveness of the proposed method.