Evaluating Negative Sampling Approaches for Neural Topic Models