Combining Statistics-Based and CNN-Based Information for Sentence Classification
Lang Zhining, Xiaozhuo Gu, Quan Zhou, Taizhong Xu · 2016
Sentence classification, serving as the foundation of the subsequent text-based processing, continues attracting researchers attentions. Recently, with the great success of deep learning, convolutional neural network (CNN), a kind of common architecture of deep learning, has been widely used to this filed and achieved excellent performance. However, most CNN-based studies focus on using complex architectures to extract more effective category information, requiring more time in training models. With the aim to get better performance with less time cost on classification, this paper proposes two simple and effective methods by fully combining information both extracted from statistics and CNN. The first method is S-SFCNN, which combines statistical features and CNN-based probabilistic features of classification to build feature vectors, and then the vectors are used to train the logistic regression classifiers. And the second method is C-SFCNN, which combines CNN-based features and statistics-based probabilistic features of classification to build feature vectors. In the two methods, the Naive Bayes log-count ratios are selected as the text statistical features and the single-layer and single channel CNN is used as our CNN architecture. The testing results executed on 7 tasks show that our methods can achieve better performance than many other complex CNN models with less time cost. In addition, we summarized the main factors influencing the performance of our methods though experiment.