Learning to Generate Equitable Text in Dialogue from Biased Training Data