Label-Embedding-Based Multi-core Convolution for Text Categorization

Text categorization is a classical problem in Natural Language Processing (NLP) tasks. There have been a number of studies that applied Convolutional Neural Networks (CNN) to this problem. This paper proposes a novel model to improve the traditional convolution for text classification, which is based on label embedding. It learns the attention on a training set oflabeled samples to ensure that the weight of related words is higher than that of irrelevant words. Meanwhile, the model also encodes the bi-gram features extracted from training data into convolutional filters, instead of randomly initializing them, which can help it focus on learning important semantic features at the beginning of the training. Experimental results on several text datasets indicate that the proposed model is competitive to the state-of-the-art models.