Cross-modality Consistent Regression for Joint Visual-Textual Sentiment Analysis of Social Multimedia