Vision-Text Cross-Modal Fusion for Accurate Video Captioning