An adaptive convolution with label embedding for text classification

Convolutional neural network (CNNs) has made a breakthrough since deep learning was employed to text classification. However, the traditional CNNs usually use the same set of filters for feature extraction, where labels play a less central role of the final performance, as label information is not better utilized. To solve the problem of how to represent label information better and how to apply the learned label representations to the text classification tasks, we propose an adaptive convolution with label embedding(ACLE) in this paper, which adaptively generated convolutional filters that are conditioned on inputs and made each label embedded in the same space with the word vectors. Our method maintains the flexibility of adaptive convolution, and fully extracts label information to play an auxiliary role. The experimental results on the several large text datasets show that the proposed model is feasible and out-performs the state-of-the-art methods by a large margin in terms of accuracy.