Modeling the interaction between sensory and affective meanings for detecting metaphor

Concreteness and imageability have long been held to play an important role in the meanings of figurative expressions. Recent work has implemented this idea in order to detect metaphors in natural language discourse. Yet, a relatively unexplored dimension of metaphor is the role of affective meanings. In this paper, we will show how combining concreteness, imageability and sentiment scores, as features at different linguistic levels, improves performance in such tasks as automatic detection of metaphor in discourse. By gradually refining these features through descriptive studies, we found the best performing classifier for our task to be random forests. Further refining of our classifiers for part-ofspeech, led to very promising results, with F1 scores of .744 for nouns,.799 for verbs, .811 for prepositions. We suggest that our approach works by capturing to some degree the complex interactions between external sensory information (concreteness), information about internal experience (imageability), and relatively subjective meanings (sentiment), in the use of metaphorical expressions in natural language.