PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7426–7441, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 7426 POWERTRANSFORMER: Unsupervised Controllable Revision for Biased Language Correction Xinyao (Michelle) Ma⋆ ⋄ Maarten Sap⋆ ⋄ Hannah Rashkin⋄ Yejin Choi⋄† ⋄Paul G. Allen School of Computer Science & Engineering, University of Washington †Allen Institute for Artificial Intelligence Seattle, USA {max36,msap,hrashkin,yejin}@cs.washington.edu Abstract Unconscious biases continue to be prevalent in modern text and media, calling for algorithms that can assist writers with bias correction. For example, a female character in a story is often portrayed as passive and powerless (“She day- dreams about being a doctor”) while a man is portrayed as more proactive and powerful (“He pursues his dream of being a doctor”). We formulate Controllable Debiasing, a new revision task that aims to rewrite a given text to correct the implicit and potentially undesirable bias in character portrayals. We then introduce POWERTRANSFORMER as an approach that debiases text through the lens of connotation frames (Sap et al., 2017), which encode prag- matic knowledge of implied power dynamics with respect to verb predicates. One key chal- lenge of our task is the lack of parallel cor- pora. To address this challenge, we adopt an unsupervised approach using auxiliary super- vision with related tasks such as paraphrasing and self-supervision based on a reconstruction loss, building on pretrained language models. Through comprehensive experiments based on automatic and human evaluations, we demon- strate that our approach outperforms ablations and existing methods from related tasks. Fur- thermore, we demonstrate the use of POWER- TRANSFORMER as a step toward mitigating the well-documented gender bias in character portrayal in movie scripts. 1 Introduction Narratives and news texts often reflect societal bi- ases and stereotypes, such as the traditional gen- der role that women are passive and submissive (Lakoff, 1973; Fiske, 1993; Fast et al., 2016). The task of controllable text revision, i.e., rephrasing text to a targeted style or framing, can help correct for these biases by altering and equalizing the way ⋆ Both authors contributed equally. Connotation Frames PowerTransformer Mey daydreams of being a doctor. agency(AG) = low to daydream AGENT Mey pursues her dream to be a doctor. agency(AG) = high to pursue AGENT PowerTransformer PowerTransformer Ana wandered through the park. Ana strutted through the park. Issa enjoyed football growing up. Issa loved playing football growing up. agency(AG) = low agency(AG) = low agency(AG) = high agency(AG) = high Figure 1: Examples of using connotation frames (Sap et al., 2017) for controllable revisions to portray charac- ters with more agency and power. In the second exam- ple, “Ana strutted” implies that she is more active and decisive, compared to “Ana wandered” which portrays her as aimless and passive. people are described. For example, automatically rewriting “Mey daydreamed about being a doctor” as “Mey pursued her dream to be a doctor” por- trays Mey with more authority and decisiveness (Figure 1). Such controllable revision methods could be used to help reshape how gender roles …

Content truncated to 500 words