Visual-Textual Sentiment Analysis Enhanced by Hierarchical Cross-Modality Interaction