Statistical Metaphor Processing for Real-world NLP Tasks

Abstract

Metaphor is highly frequent in language, which makes its computational processing indispensablefor real-world NLP applications addressing semantic tasks. Previous approaches tometaphor modelling rely on task-specific hand-coded knowledge and operate on a limited domainor a subset of phenomena. We present the first integrated open-domain statistical model ofmetaphor processing in unrestricted text. Our method first identifies metaphorical expressionsin running text and then paraphrases them with their literal paraphrases. Such a text-to-textmodel of metaphor interpretation is compatible with other NLP applications that can benefit frommetaphor resolution. Our approach relies on the state-of-the-art parsing and lexical acquisitiontechnologies (distributional clustering and selectional preference induction) and operates with ahigh accuracy. Since it is minimally supervised, it can be easily ported across domains and tasks.