A Comparison of Algorithms for Detection of 'Figurativeness' in Metaphor, Irony and Puns

摘要

Figurative speech is an umbrella term for metaphor, irony, sarcasm, puns and some other speech genres and figures of speech. In research and competitions like SemEval, each of them is usually processed separately with a task-specific model. However, being altogether called "figurative speech", they should share some property: "figurativeness". If such a property exists, figurative speech can be processed simultaneously by one and the same algorithm. The present research compares performance of several NLP methods that were designed to detect one type of figurative speech (either metaphor, or irony, or puns) on short texts containing a combination of these types. The study shows that, despite being task-specific, state-of-the-art algorithms are able to process different types of figurative speech fairly well, and some of them are good even at cross-detection when the training set contains one type and the test set another.