In this paper, convolutional neural network architecture that jointly uses features for word and character embedding is proposed. This is the first paper that deals with the text classification task with word and character embedding feeding into the ConvNet at the same time. Our model uses independent sets of filters for word2vec embedding and one hot vector character-level embedding of a text, merges extracted features at fully connected layers to classify the text.
It is shown through series of text classification experiments that the proposed architecture can outpeform other models which adopt only one form of embedding such as word or character. Our model also converged 2 times faster than the other model which uses only characters.