Quantifying and Reducing Stereotypes in Word Embeddings