Improving language models with novel contrastive learning objectives