A Multi-View Semantic Autoencoder Algorithm for Human-Centric Text Sentiment Analysis

In the artificial intelligence-empowered Internet of Things (IoT), the diversification of data collection makes the collected text data to be more complex together with the diversified characteristics of multi-description. Technically, such data are generally referred to multi-view data in text sentiment. How to mine and utilize the complementary and consistent knowledge to perform sentiment classification for text data is a core issue to be solved in multi-view learning. In this study, a multi-view semantic autoencoder (MVSA) algorithm is proposed based on semantic autoencoder model and label relaxation strategy. MVSA algorithm learns the mapping function from sample space to semantic space for maintaining the consistency of each view. This semantic space not only contains the discriminative information of text data, but also contains the information of the class relationships in the semantic space. The relaxation strategy and low-rank constraint on the label matrix ensure the compactness of predicted label and inter-class separation as large as possible. Finally, the test sample can be classified by various classifiers. The experimental results show the MVSA algorithm achieves high classification performance, which proves the effectiveness of the proposed algorithm in text sentiment analysis.