Towards Understanding Gender Bias in Relation Extraction

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2943–2953 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 2943 Towards Understanding Gender Bias in Neural Relation Extraction Andrew Gaut*†, Tony Sun*†, Shirlyn Tang†, Yuxin Huang†, Jing Qian†, Mai ElSherief††, Jieyu Zhao‡, Diba Mirza†, Elizabeth Belding†, Kai-Wei Chang‡, and William Yang Wang† †Department of Computer Science, UC Santa Barbara ‡Department of Computer Science, UC Los Angeles ††School of Interactive Computing, Georgia Institute of Technology {ajg, tonysun, shirlyntang, yuxinhuang}@ucsb.edu {jing qian, dimirza, ebelding, william}@cs.ucsb.edu melsherief@gatech.edu {jyzhao, kwchang}@cs.ucla.edu Abstract Recent developments in Neural Relation Ex- traction (NRE) have made significant strides towards automated knowledge base construc- tion. While much attention has been dedicated towards improvements in accuracy, there have been no attempts in the literature to evaluate social biases exhibited in NRE systems. In this paper, we create WikiGenderBias, a distantly supervised dataset composed of over 45,000 sentences including a 10% human annotated test set for the purpose of analyzing gender bias in relation extraction systems. We find that when extracting spouse and hypernym (i.e., occupation) relations, an NRE system performs differently when the gender of the target entity is different. However, such dispar- ity does not appear when extracting relations such as birth date or birth place. We also an- alyze two existing bias mitigation techniques, word embedding debiasing and data augmenta- tion. Unfortunately, due to NRE models rely- ing heavily on surface level cues, we find that existing bias mitigation approaches have a neg- ative effect on NRE. Our analysis lays ground- work for future quantifying and mitigating bias in relation extraction. 1 Introduction With the wealth of information being posted on- line daily, relation extraction has become increas- ingly important. Relation extraction aims specifi- cally to extract relations from raw sentences and represent them as succinct relation tuples of the form (head, relation, tail) e.g., (Barack Obama, spouse, Michelle Obama). * Equal Contribution. The concise representations provided by relation extraction models have been used to extend Knowl- edge Bases (KBs) (Riedel et al., 2013; Subasic et al., 2019; Trisedya et al., 2019). These KBs are then used heavily in NLP systems, such as question answering systems (Bordes et al., 2014; Yin et al., 2016; Cui et al., 2019). In recent years, much focus in the Neural Relation Extraction (NRE) commu- nity has been centered on improvements in model precision and the reduction of noise (Lin et al., 2016; Liu et al., 2017; Wu et al., 2017; Feng et al., 2018; Vashishth et al., 2018; Qin et al., 2018). Yet, little attention has been devoted towards the fair- ness of such systems. We take the first step at understanding and evalu- ating gender bias in NRE systems by measuring the differences in model performance when extracting relations from sentences written about females ver- sus sentences written about males. If a NRE model predicts a relation such occupation with higher re- call on male entities, this could lead to the resulted knowledge bases having more occupation …

Content truncated to 500 words