Fast Korean Syllable Recognition with Letter-Based Convolutional Neural Networks

Abstract

Convolutional neural networks are widely used for character recognition. However, a large number of characters (that is, a number of classes) in CJK (Chinese, Japanese, Korean) languages can lead to problems in a straightforward approach. Also the fact that the distribution of the characters in these languages is far from a uniform distribution can cause additional problems. For example, about 10% of Korean syllables cover 99.9% of Korean texts, so image search of rare syllables for a training set can be challenging. In this paper we introduce a letter-based Korean syllable recognition approach which is able to resolve the problems described above. The core idea of the approach is to recognize the letters that make up a syllable separately instead of recognizing the whole syllable. Additional advantages of the proposed approach are high computational speed and small size of the model. The paper also briefly describes model deployment.