Adjusting Interpretable Dimensions in Embedding Space with Human Judgments