Learning Visual-Semantic Embeddings for Reporting Abnormal Findings on Chest X-rays