Gender classification of Korean personal names: Deep neural networks versus human judgments

Abstract

In many languages, female and male names have different phonotactic characteristics. The name–gender relationship is probabilistic; therefore, it can be captured more adequately using stochastic models than deterministic phonological theories. In this study, a total of 6,000 most commonly used names (3,000 for each gender) in Korean were used to train a deep neural network (DNN), which is an ensemble model of recurrent neural networks and convolution neural networks. The phonotactic learner (PL) was used as the baseline model. The DNN and PL models predicted the gender of 50 test names compiled from low-frequency names. The models’ predictions were compared with human judgments on the gender of the test names. The models’ predicted labels matched the names’ actual labels, with a higher accuracy in the DNN (90%) than in the PL (76%). The predictions also matched the labels assigned by human subjects with a higher accuracy for the DNN (86%) than the PL (72%). The DNN model correlated more closely with human judgments (r² = 0.743) than the PL (r² = 0.312). Considering the similarity of responses between the DNN and humans, these results suggest that neural network models should be incorporated into phonological studies.