Self-Supervised Document Similarity Ranking via Contextualized Language Models and Hierarchical Inference