Undesirable biases in NLP: Averting a crisis of measurement