On the role of shape features in metaphor generation for abstract images

In this study, we examined the role of shape features on metaphor generation for abstract images based on a simulation with a retrained convolutional neural network (CNN), which is in turn based on a pretrained CNN model (AlexNet). A computational experiment was conducted using three types of object recognition models, including a pretrained object recognition model (AlexNet) and recognition models that were retrained to recognize more or fewer shape features using edge-detected or blurred images from the ILSVRC-2012 dataset. A psychological experiment was conducted to collect metaphors that were used to explain the abstract images. The simulation results of the models for the abstract images were compared to examine how well they predicted the concepts used in the metaphors generated for the abstract images. The results of the computational experiment suggest that the model retrained to recognize fewer shape features performed best at predicting the generated metaphors. However, for some abstract images, the model retrained to recognize more shape features performed better. These results suggest that the role of shape features on metaphor generation differs depending on the types of abstract images.