Knowledge graph-based metaphor representation for literature understanding

Abstract

Metaphor understanding is one of the most important and challenging tasks in language processing. In this study, we propose a knowledge graph-based metaphor embedding for understanding metaphor relations between words in literature. We consider that there is a metaphor relation between two objects involved in metaphorical rhetoric (e.g. in the metaphor 'Time is money', 'time', and 'money' have a metaphor relation). We first propose extracting metaphor relations between words based on part of speech (POS) tagging, then a knowledge graph-based metaphor representation method is presented. The proposed approach enables (i) metaphor interpretation in literature based on the proposal of extracting metaphor relations and (ii) representation of metaphor based on metaphor graph that consists of the nodes of words and the edges of their metaphor relations. Finally, we apply our metaphor embedding to Shakespeare's sonnets for evaluating the performance of it. Experimental results of predicting metaphor relations and classifying metaphorical triples prove that our approach achieves creative results. Besides, the literary significance of our approach can be embodied by the analysis of metaphors in Shakespeare's sonnets.