Learning from corpus and learning from supervised NLP tasks both give useful semantics that can be incorporated into a good word representation. We propose an embedding learning method called Delta Embedding Learning, to learn semantic information from high-level supervised tasks like reading comprehension, and combine it with an unsupervised word embedding. The simple technique not only improved the performance of various supervised NLP tasks, but also simultaneously learns improved universal word embeddings out of these tasks.