Comparative Analysis on Encoding Methods for Convolutional Neural Network in Korean Text Classification

Abstract

Text classification is an essential task for natural language processing. Among those classifiers, Convolutional Neural Network(CNN) have recently shown strong performance in text classification. However, those researches are based on words and word-level CNN requires vast word vectors and morphological analyzers in Korean language. This study excluded the use of morphological analyzers and compared the results of classifying Korean internet news articles among different input levels of CNN. The experiment result shows that syllable-level CNN performs as well as word-level CNN, while character-level CNN shows weak performance.