Monolingual and Multilingual Reduction of Gender Bias in Contextualized Representations

Proceedings of the 28th International Conference on Computational Linguistics, pages 5082–5093 Barcelona, Spain (Online), December 8-13, 2020 5082 Monolingual and Multilingual Reduction of Gender Bias in Contextualized Representations Sheng Liang, Philipp Dufter, Hinrich Sch¨utze Center for Information and Language Processing (CIS) LMU Munich, Germany {shengliang, philipp}@cis.lmu.de Abstract Pretrained language models (PLMs) learn stereotypes held by humans and reflected in text from their training corpora, including gender bias. When PLMs are used for downstream tasks such as picking candidates for a job, people’s lives can be negatively affected by these learned stereotypes. Prior work usually identifies a linear gender subspace and removes gender information by eliminating the subspace. Following this line of work, we propose to use DensRay, an analytical method for obtaining interpretable dense subspaces. We show that DensRay performs on-par with prior approaches, but provide arguments that it is more robust and provide indications that it preserves language model performance better. By applying DensRay to attention heads and layers of BERT we show that gender information is spread across all attention heads and most of the layers. Also we show that DensRay can obtain gender bias scores on both token and sentence levels. Finally, we demonstrate that we can remove bias multilingually, e.g., from Chinese, using only English training data. 1 Introduction [MASK] kissed [MASK] . The professor asked the nurse . The child played with the car . The child played with the doll . Table 1: Token level bias scores (male/female bias depicted by blue/red) identified using DensRay. See Table 6 for more details. Word embeddings, which represent the semantic meaning of text data as vectors, are used as input in natural language pro- cessing tasks. It has been found that word embeddings exhibit biases such as gender bias, which are present in their training corpora (Bolukbasi et al., 2016; Caliskan et al., 2017; Garg et al., 2018). Contextual word embedding models, such as BERT (Devlin et al., 2019), have become increasingly com- mon and achieved new state-of-the-art results in many NLP tasks. Researchers have also found gender bias in contextu- alized embeddings (Zhao et al., 2019; May et al., 2019). It has been found that these biases have a big impact on downstream tasks (Vanmassenhove et al., 2018; Moryossef et al., 2019; Rudinger et al., 2018; Zhao et al., 2018). A common approach for removing gender information in static embeddings is to identify a linear gender subspace (e.g., a gender direction) and subsequently setting all values on the gender direction to zero. Successful approaches rely on simple principal component analysis (Bolukbasi et al., 2016; Mu and Viswanath, 2018). Bolukbasi et al. (2016) require pairs of gendered words to compute a direction (e.g., “man”-“woman”) and Mu and Viswanath (2018) rely on computing a PCA of a set of gender words hoping that the main variation direction is the gender direction. We propose to use DensRay (Dufter and Sch¨utze, 2019) which only requires two or multiple groups of gendered words. In summary, our contributions are: i) We adjust …

Content truncated to 500 words