Attention Correctness in Neural Image Captioning