Topic discovery via latent space clustering of pretrained language model representations