What makes a metaphor literary? Answers from two computational studies

ABSTRACT In this article we investigate structural differences between “literary” metaphors created by renowned poets and “nonliterary” ones imagined by non-professional authors from Katz et al.’s 1988 corpus. We provide data from quantitative narrative analyses (QNA) of the altogether 464 metaphors on over 70 variables, including surface features like metaphor length, phonological features like sonority score, or syntactic-semantic features like sentence similarity. In a first computational study using machine learning tools (i.e., a classifier of the decision tree family) we show that Katz et al.’s literary metaphors can be successfully discriminated from their nonliterary ones on the basis of response measures (10 ratings), in particular the ratings for familiarity, ease of interpretation, semantic relatedness, and comprehensibility. A second computational study then shows that the classifier can reliably detect and predict between-group differences on the basis of five QNA features generalizing from a training to a test corpus. Our results shed light on surface and semantic features that co-determine the reception of metaphors and raise important questions about their literariness, aptness or poetic potential. They tentatively suggest a set of 11 features that could influence the “literariness” of metaphors, including their sonority score, length and surprisal value.