Evaluating Unsupervised Text Classification: Zero-shot and Similarity-based Approaches