Queens are Powerful too: Mitigating Gender Bias in Dialogue Generation

Abstract

Social biases present in data are often directly reflected in the predictions of models trained on that data. We analyze gender bias in dialogue data, and examine how this bias is not only replicated, but is also amplified in subsequent generative chit-chat dialogue models. We measure gender bias in six existing dialogue datasets before selecting the most biased one, the multi-player textbased fantasy adventure dataset LIGHT (Urbanek et al., 2019), as a testbed for bias mitigation techniques. We consider three techniques to mitigate gender bias: counterfactual data augmentation, targeted data collection, and bias controlled training. We show that our proposed techniques mitigate gender bias by balancing the genderedness of generated dialogue utterances, and find that they are particularly effective in combination. We evaluate model performance with a variety of quantitative methods-including the quantity of gendered words, a dialogue safety classifier, and human assessments-all of which show that our models generate less gendered, but equally engaging chit-chat responses.

Multi-Dimensional Gender Bias Classification

“…For dialogue, gender biases in training corpora have been found to be amplified in machine learning models Dinan et al, 2020;Liu et al, 2019). While many of the works cited above proposed methods of mitigating the unwanted effects of gender on text, Hall Maudslay et al ( 2019), Liu et al (2019), Zmigrod et al (2019), andDinan et al (2020) in particular relied on counterfactual data to alter the training distribution to offset gender-based statistical imbalances (see §4.2 for more discussion of training set imbalances). Also relevant is Kang et al (2019, PASTEL), which introduced a parallel style corpus and showed gains on style-transfer across binary genders.…”

“…By learning to associate control variables with textual properties, generative models can be controlled at inference time to adjust the generated text based on the desired properties of the user. This has been applied to a variety of different cases, including generating text of different lengths (Fan et al, 2018a), generating questions in chit-chat (See et al, 2019), and reducing bias (Dinan et al, 2020).…”

“…ited in coverage and applicability to a variety of domains (Dinan et al, 2020). However, by decomposing bias along the TO, AS, and ABOUT dimensions, fine-grained control models can be trained to control these different dimensions separately.…”

“…We also compare to a baseline for which the control tokens are determined by a word list: if an utterance contains more masculine-gendered words than feminine-gendered words from the word list it is labeled as masculine (and vice versa for feminine); if it contains no gendered words or an equal number of masculine and feminine gendered words, it is labeled as neutral. Following Dinan et al (2020), we combine several existing word lists (Zhao et al, 2018bHoyle et al, 2019).…”

“…In particular, models often can unwittingly learn negative associations about protected groups present in their training data and propagate them. In particular, NLP models often learn to replicate unwanted gender biases present in society (Bolukbasi et al, 2016;Hovy and Spruit, 2016;Caliskan et al, 2017;Rudinger et al, 2017;Garg et al, 2018;Dinan et al, 2020). Since unwanted gender biases can affect downstream applications-sometimes even * Joint first authors.…”