Rethinking Content and Style: Exploring Bias for Unsupervised Disentanglement