Heterogeneous Transfer with Deep Latent Correlation for Sentiment Analysis

Most traditional methods of image sentiment analysis focus on the design of visual features, and the usefulness of texts associated to the images have not been sufficiently investigated. Heterogeneous transfer learning has recently gained much attention as a new machine learning paradigm in which knowledge can be transferred from source domain feature space to target domain feature space. This paper proposes a novel approach that exploits deep latent correlation between visual and textual modalities. In our proposed method, we build a latent embedding space for symmetric heterogeneous feature transfer. The latent space is able to generate domain-specific and maximally correlative cross-domain features which are regarded as the semantic-intensive visual feature representation and used to train sentiment polarity classifiers. The results of experiments conducted on real-world data sets show that the proposed approach can achieve better sentiment classification accuracy by using multi-layer neural network to capture deeper internal relations.