Improving Sentence Embeddings with an Automatically Generated NLI Dataset