A Character-level Convolutional Neural Network (Char-CNN) is an efficient method for text categorization. This method uses an input from characters, therefore, when applying it to categorize Thai text, a word segmentation step is not required. However, an original model of Char-CNN limits an input length to 1,014 characters. Any exceeding character is ignored. This thesis presents an improvement of Char-CNN which can accept any input length while it still uses the same number of parameters. Experiments show that our proposed model can produce a better accuracy than an original model. Moreover, the proposed technique outperforms many classical techniques e.g. Naïve Bayes, Maximum Entropy and Support Vector Machine. Note that there is only one technique, a word-level Convolutional Neural Network, that it performs better than our model about 0.5%. However, a Char-CNN has an advantage because its accuracy does not depend on a performance of word segmentation.