Understanding Gender Bias in Knowledge Base Embeddings

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1381 - 1395 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Understanding Gender Bias in Knowledge Base Embeddings Yupei Du♠ Qi Zheng♥ Yuanbin Wu♥ Man Lan♥ Yan Yang♥♦ Meirong Ma♣ ♠Department of Information and Computing Sciences, Utrecht University, the Netherlands ♥Department of Computer Science and Technology, East China Normal University, China ♦Shanghai Key Laboratory of Multidimensional Information Processing, China ♣Shanghai Transsion Co., Ltd, China y.du@uu.nl, qizheng.ecnu@outlook.com {ybwu,mlan,yanyang}@cs.ecnu.edu.cn, meirong.ma@transsion.com Abstract Knowledge base (KB) embeddings have been shown to contain gender biases (Fisher et al., 2020b). In this paper, we study two ques- tions regarding these biases: how to quantify them, and how to trace their origins in KB? Specifically, first, we develop two novel bias measures respectively for a group of person entities and an individual person entity. Evi- dence of their validity is observed by compar- ison with real-world census data. Second, we use influence function to inspect the contribu- tion of each triple in KB to the overall group bias. To exemplify the potential applications of our study, we also present two strategies (by adding and removing KB triples) to mitigate gender biases in KB embeddings. 1 Introduction Gender biases have been shown to have noticeable presence in a wide range of NLP models. For ex- ample, we can observe that the word embedding of “engineer” is closer to “he” than “she” (Bolukbasi et al., 2016), and co-reference systems associate “surgeon” more with masculine pronouns than with feminine ones (Rudinger et al., 2018). These biases are brought to our models from training data by our algorithms. Hence, besides revealing the existence of gender biases, it is important to quantify them and explain their origins in data. Knowledge bases (KB, e.g. Freebase, Bollacker et al., 2007) provide accessible organizations of human knowledge by the form of triples. Each triple consists of a head entity, a relation, and a tail entity. For example, the fact that Marie Curie is a chemist is represented as ⟨Marie Curie, people.person.profession, chemist⟩. KB embeddings encode these knowledge into dense vector representations. It is important to understand gender biases in KB embeddings for two major reasons. First, KB embeddings serve as sources of prior knowledge in many downstream NLP models (e.g. pre-trained language models, Zhang et al., 2019). Clearly, if biases exist in KB embeddings, they can easily propagate into these models, and drive these models more biased. Sec- ond, Garg et al. (2018) observe that word embed- dings reflect biases in the training corpora, and hence our society. Likewise, we suspect KB em- beddings to reflect biases encoded in KBs, as also suggested by Radstok et al. (2021). In this paper, we propose two novel gender bias measures for KB embeddings, one for a group of person entities (group bias) and the other for an individual person entity (individual bias). Fur- thermore, with influence function (Koh and Liang, 2017), we explain the origins of group bias at the …

Content truncated to 500 words