Improving Contextualized Topic Models with Negative Sampling