A character-level convolutional neural network with dynamic input length for Thai text categorization
Thanabhat Koomsubha, Peerapon Vateekul · 2017
A Character-level Convolutional Neural Network (Char-CNN) is an efficient text categorization method. It can be used in categorization task without a word segmentation step, which is necessary by traditional method for Thai. Currently, the existing model of Char-CNN uses a fixed input length and requires cutting off exceeding characters, which may lead to a missing of important content. In this paper, we propose a new Char-CNN model with a capability to accept any length of input by employing k-max pooling before a fully connected layer. The result shows that our model outperforms a Char-CNN model with a fixed input length on Thai news categorization. Moreover, our proposed method gives a better accuracy than many word-level methods: Naive Bayes, Logistic Regression, Support Vector Machine except a word-level CNN.