This is HAN version of sentiment classification, with pre-trained character-level embedding, and used RAN instead of GRU.
Downloaded from internet but i forget where is it ;p, the original dataset is in Simplified Chinese, i used opencc translated it into Traditional Chinese. After 100 epochs, the valid accuracy achieved 96.31%
Tensorflow r1.1+