PoEM : a parser of emotion metaphors

Abstract

Although metaphor is generally recognized as an integral component of everyday language, very few computational systems capable of understanding metaphoric utterances exist today. This thesis describes one approach to the problem and presents PoEM, a prototype system which recognizes and interprets metaphoric descriptions of emotions and mental states in single-sentence input. Building upon previous work in knowledge-based metaphor comprehension, this research adopts a goal-driven approach which assumes each metaphor is selected by a speaker for its aptness at serving a particular communicative goal. To identify these goals, an empirical analysis of metaphor distribution in song lyrics was performed, and typical communicative intentions and surface patterns were identified for the top five most frequently occurring metaphor groups. These intentions and surface patterns have been implemented as a set of metaphor templates and interpretation rules in PoEM, using the WordNet lexical database for supplemental semantic information. Evaluation of PoEM demonstrates fairly high accuracy but low recall.