Incorporating GAN for Negative Sampling in Knowledge Representation Learning