Co-training, a paradigm of semi-supervised learning, may alleviate effectively the data scarcity problem (i.e., the lack of labeled examples) in supervised learning. The standard two-view co-training requires the dataset be described by two views of attributes, and previous theoretical studies proved that if the two views satisfy the sufficiency and independence assumptions, co-training is guaranteed to work well. However, little work has been done on how these assumptions can be empirically verified given datasets. In this paper, we first propose novel approaches to verify empirically the two assumptions of co-training based on datasets. We then propose simple heuristic to split a single view of attributes into two views, and discover regularity on the sufficiency and independence thresholds for the standard two-view co-training to work well. Our empirical results not only coincide well with the previous theoretical findings, but also provide a practical guideline to decide when co-training should work well based on datasets.