“…My husband is a good man, but only seems to stay with me out of duty. ing and mitigating gender bias for three additional reasons: first, it has multiple potential sources of bias-not just dialogues, but also characters and personas-second, it was crowdsourced, and thus susceptible to reflecting the gender biases known to be present in crowdworkers' annotations (Otterbacher et al, 2018;Barbosa and Chen, 2019), and third, LIGHT's medieval, fantasy setting might encourage crowdworkers to impart text with their gender biases. After selecting LIGHT for particular scrutiny, we then explore three bias mitigation techniques, one of which is wholly novel, and another which is novel in its application to dialogue: (i) Counterfactual Data Augmentation (CDA) (Hall Maudslay et al, 2019;Zmigrod et al, 2019), (ii) a targeted data collection method, which we refer to as Positive-Bias Data collection, and (iii) Bias Controlled text generation.…”