RAN: Radical analysis networks for zero-shot learning of Chinese characters
Chinese characters have a huge set of character categories, more than 20,000 and the number is still increasing as more and more novel characters continue being created. However, the enormous characters can be decomposed into a few fundamental structural radicals, only about 500. This paper introduces the Radical Analysis Networks (RAN) that recognize Chinese characters by identifying radicals and analyzing 2D spatial structures between them. The proposed RAN first extracts visual features from Chinese character input by employing convolutional neural networks as an encoder. Then a decoder based on recurrent neural networks is employed, who aims to generate a caption of Chinese character by detecting radicals and 2D structures through a spatial attention mechanism. The manner of treating Chinese character input as a composition of radicals rather than a single picture severely reduces the size of vocabulary and enables RAN to possess the ability of recognizing unseen Chinese character classes only if their radicals have been seen, called zero-shot learning. We test a simple implementation of RAN on experiments of recognizing printed Chinese characters with seen and unseen classes and RAN simultaneously obtains convincing performance on unseen classes and state-of-the-art performance on seen classes.
READ FULL TEXT