Understanding Metaphors Using Emotions

Abstract

Metaphors convey unspoken emotions and perceptions by creatively applying an evocative concept from the source domain to illustrate some latent idea in the target domain. Prior research on nominal metaphor interpretation focused on identifying those properties of the source domain which are highly related to the target domain concepts to discover the most likely sense of the metaphor’s usage. In this paper, we bring forth a fresh perspective by observing that a metaphor is seldom without an emotion or sentiment; in fact, it is this very aspect which segregates it from its literal counterpart. We present an Emotion driven Metaphor Understanding system which assesses the affective dimensions of the source properties before assigning them as the most plausible sense in the context of the target domain. In our approach, we use the web as a knowledge source to identify properties of the source domain. We resolve the bottleneck of non-availability of informative emotion lexicons using pre-trained word2vec embeddings to extract the latent emotions in the source domain properties. Adopting an unsupervised learning approach on a dataset of nominal metaphors, we demonstrate that in comparison with a single emotionless interpretation, a multi-sense interpretation of a metaphor using the gamut of emotions is more likely to provide a realistic presentation of its purport. We further demonstrate that an emotion driven interpretation is often preferred over an interpretation sans emotion. The results clearly indicate that it is beneficial to apply emotions for refining the process of metaphor understanding.