Queens are Powerful too: Mitigating Gender Bias in Dialogue Generation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 8173–8188, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 8173 Queens are Powerful too: Mitigating Gender Bias in Dialogue Generation Emily Dinan∗, Angela Fan∗†, Adina Williams, Jack Urbanek, Douwe Kiela, Jason Weston Facebook AI Research †Laboratoire Lorrain d’Informatique et Applications (LORIA) Abstract Social biases present in data are often di- rectly reflected in the predictions of mod- els trained on that data. We analyze gender bias in dialogue data, and examine how this bias is not only replicated, but is also am- plified in subsequent generative chit-chat di- alogue models. We measure gender bias in six existing dialogue datasets before select- ing the most biased one, the multi-player text- based fantasy adventure dataset LIGHT (Ur- banek et al., 2019), as a testbed for bias mit- igation techniques. We consider three tech- niques 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 dia- logue utterances, and find that they are partic- ularly effective in combination. We evaluate model performance with a variety of quantita- tive methods—including the quantity of gen- dered words, a dialogue safety classifier, and human assessments—all of which show that our models generate less gendered, but equally engaging chit-chat responses. 1 Introduction Machine learning algorithms learn to model pat- terns present in training datasets. In particu- lar, they make predictions that directly reflect the harmful societal biases present in training datasets, such as racial bias in sports reports (Merullo et al., 2019) and political bias in news data (Fan et al., 2019). Such biases are rife in NLP, for exam- ple, in learned word embeddings (Bolukbasi et al., 2016; Brunet et al., 2018; Zhao et al., 2019), vi- sual semantic role labeling (Zhao et al., 2017), natural language inference (He et al., 2019), abu- sive language classification (Park et al., 2018), and ∗Joint first authors. Gendered word counts in dialogue datasets Dataset % gend. words % male bias LIGHT 0.94 73.4 Reddit 1.32 69.76 Wizard of Wikipedia 0.076 65.9 Daily Dialog 1.02 59.04 Empathetic Dialogues 2.07 53.45 ConvAI2 1.28 50.05 Table 1: Counts of gendered words in several di- alogue datasets. We report the percent of gendered words (% gend. words) as well as the percentage of male-gendered words out of all gendered words (% male bias). Datasets are arranged in descending order with respect to % male bias. LIGHT has the most % male bias; thus we chose it as our main testbed. coreference resolution (Zhao et al., 2018a). Al- though research into bias in NLP writ large is ma- turing, bias in dialogue utterances has received somewhat less attention (Liu et al., 2019; Sheng et al., 2019; Henderson et al., 2018). As real- world use-cases for dialogue agents, such as in- teractive assistants, are rapidly developing, bias in dialogue models has the very real potential to in- vade downstream systems and exacerbate existing …

Content truncated to 500 words