Topic modeling techniques have been widely applied in many cloud computing applications. However, few of them have tried to discover latent semantic relationships of implicit topics and explicit words to generate a more comprehensive representation for each text. To fully exploit the semantic knowledge for text classification in cloud computing systems, we attempt to encode topic and word features based on their latent relationships. The extracted topical information reorganizes the original textual structures from two aspects: one is that the topic extracted by Latent Dirichlet Allocation (LDA) is viewed as a textual extension; the other is that the topic feature performs as a counterpart modality to the word. This paper proposes a Dual Semantic Embedding (DSE) method, which uses Convolutional Neural Networks (CNNs) to encode the dual semantic features of topics and words from the reorganized semantic structures. Experimental results show that DSE improves the performance of text classification and outperforms the state-of-the-art feature generation baselines on micro- F1 and macro- F1 scores over the real-world text classification datasets.
