Adaptive-Rank Fusion For Sentiment Analysis: End-To-End Training And Generalization Behavior