Assessing the Reliability of Word Embedding Gender Bias Measures

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10012–10034 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 10012 Assessing the Reliability of Word Embedding Gender Bias Measures Yupei Du, Qixiang Fang and Dong Nguyen Utrecht University Utrecht, the Netherlands {y.du,q.fang,d.p.nguyen}@uu.nl Abstract Various measures have been proposed to quan- tify human-like social biases in word embed- dings. However, bias scores based on these measures can suffer from measurement error. One indication of measurement quality is reli- ability, concerning the extent to which a mea- sure produces consistent results. In this pa- per, we assess three types of reliability of word embedding gender bias measures, namely test- retest reliability, inter-rater consistency and in- ternal consistency. Specifically, we investigate the consistency of bias scores across different choices of random seeds, scoring rules and words. Furthermore, we analyse the effects of various factors on these measures’ reliability scores. Our findings inform better design of word embedding gender bias measures. More- over, we urge researchers to be more critical about the application of such measures.1 1 Introduction Despite their success in various applications, word embeddings have been shown to exhibit a range of human-like social biases. For example, Caliskan et al. (2017) find that both GloVe (Pennington et al., 2014) and skip-gram (Mikolov et al., 2013) embed- dings associate pleasant terms (e.g. love and peace) more with European-American names than with African-American names, and that they associate career words (e.g. profession and business) more with male names than with female names. Various measures have been proposed to quan- tify such biases in word embeddings (Ethayarajh et al., 2019; Zhou et al., 2019; Manzini et al., 2019). These measures allow us to assess biases in word embeddings and the performance of bias-mitigation methods (Bolukbasi et al., 2016; Zhao et al., 2018). They also enable us to study social biases in a new way complementary to traditional qualitative 1Our code is available at https://github.com/ nlpsoc/reliability_bias. fj (e.g. woman, mother) w1 poetry w2 literature w3 drama mj (e.g. man, father） query q wi symphony ... gender base pair (m, f) target word w 𝑤𝑤2 𝑤𝑤3 𝑤𝑤i 𝑚𝑚𝑗𝑗 ⃗𝑓𝑓𝑗𝑗 𝑤𝑤1 Figure 1: Measuring word embedding gender biases in the concept of arts. The arrows are hypothetical low- dimensional projections of word embeddings. and experimental approaches (Garg et al., 2018; Chaloner and Maldonado, 2019). A key challenge in developing bias measures is that social biases are abstract concepts that can- not be measured directly but have to be inferred from some observable data. This renders the re- sulting bias scores more prone to measurement errors. Therefore, it is important to carefully assess these measures’ measurement quality. Jacobs and Wallach (2021) also highlight similar measurement issues in the context of fairness research. In this paper, we focus on one aspect of measure- ment quality: reliability. It concerns the extent to which a measure produces consistent results. In particular, we investigate the reliability of word embedding gender bias measures. Figure 1 illustrates how gender biases in word …

Content truncated to 500 words