A short text topic classification method based on feature expansion and bi-directional neural network

Short text classification has a pivotal role in many applications such as sentiment analysis and recommender systems. However, the traditional classification models work poorly on short texts due to data sparsity issue. The objective of this research is to determine whether fused topic features will extend short text features and achieve better classification performance. We propose a novel short text topic classification approach and put the topic features into consideration. Firstly, LDA topic model is used to obtain short text topic features. Next, we use BiLSTM(Bi-directional Long Short-Term Memory), CNN(Convolutional Neural Networks) and attention mechanism to extended short text feature extraction. Finally, a SoftMax classifier is introduced to obtain the topic classification results of the short texts. Compared with the state-of-the-art short text topic classification approaches, the experimental results demonstrate that our method achieves better performance.