What makes a good metaphor? A cross-cultural study of computer-generated metaphor appreciation

Computers are now able to automatically generate metaphors, but some automatically-generated metaphors are more well-received than others. In this paper, we showed participants a series of ‘A is B’ type metaphors that were either generated by humans or taken from the Twitter account ‘@Metaphorismybusiness’, which is linked to a fully automated metaphor generator. We used these metaphors to assess linguistic factors that drive metaphor appreciation and understanding, including the role of novelty, word frequency, concreteness and emotional valence of the topic and vehicle terms. We additionally assessed how these metaphors were understood in three languages, including English, Spanish and Mandarin Chinese, and whether participants thought they had been generated by a human or a computer. We found that meaningfulness, appreciation, speed in finding meaning and humanness ratings were reliably correlated with each other in all three languages, which we interpret to indicate a more general property of ‘metaphor quality’. We furthermore found that in all three languages, conventional metaphors and those that contained an ‘optimal’ (intermediate) degree of novelty were more likely to be perceived to be of higher quality than those that were extremely creative. Further analysis of the English data alone revealed that those metaphors that contained negatively valenced vehicle words and infrequent vehicle terms (in comparison with the topic terms) were more likely to be considered high-quality metaphors. We discuss the implications of these findings for the (improvement of) automatic generation of metaphor by computers, for the persuasive function of metaphor, and for theories of metaphor understanding more generally.