Learning Bias-reduced Word Embeddings Using Dictionary Definitions