Multilingual Twitter Corpus and Baselines for Evaluating Demographic Bias in Hate Speech Recognition

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1440–1448 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1440 Multilingual Twitter Corpus and Baselines for Evaluating Demographic Bias in Hate Speech Recognition Xiaolei Huang1∗, Linzi Xing2, Franck Dernoncourt3, Michael J. Paul1 1.University of Colorado Boulder, 2. University of British Columbia 3. Adobe Research 1. {xiaolei.huang, mpaul}@colorado.edu, 2. lzxing@cs.ubc.ca 3. dernonco@adobe.com Abstract Existing research on fairness evaluation of document classification models mainly uses synthetic monolingual data without ground truth for author demographic attributes. In this work, we assemble and publish a multilingual Twitter corpus for the task of hate speech detection with inferred four author demographic factors: age, country, gender and race/ethnicity. The corpus covers five languages: English, Italian, Polish, Portuguese and Spanish. We evaluate the inferred demographic labels with a crowdsourcing platform, Figure Eight. To examine factors that can cause biases, we take an empirical analysis of demographic predictability on the English corpus. We measure the performance of four popular document classifiers and evaluate the fairness and bias of the baseline classifiers on the author-level demographic attributes. Keywords: demographic bias, fairness, multilingual, document classification, hate speech 1. Introduction While document classification models should be objective and independent from human biases in documents, research have shown that the models can learn human biases and therefore be discriminatory towards particular demographic groups (Dixon et al., 2018; Borkan et al., 2019; Sun et al., 2019b). The goal of fairness-aware document classifiers is to train and build non-discriminatory models towards peo- ple no matter what their demographic attributes are, such as gender and ethnicity. Existing research (Dixon et al., 2018; Kiritchenko and Mohammad, 2018; Park et al., 2018; Garg et al., 2019; Borkan et al., 2019) in evaluating fairness of document classifiers focus on the group fairness (Choulde- chova and Roth, 2018), which refers to every demographic group has equal probability of being assigned to the positive predicted document category. However, the lack of original author demographic attributes and multilingual corpora bring challenges towards the fair- ness evaluation of document classifiers. First, the datasets commonly used to build and evaluate the fairness of doc- ument classifiers obtain derived synthetic author demo- graphic attributes instead of the original author information. The common data sources either derive from Wikipedia toxic comments (Dixon et al., 2018; Park et al., 2018; Garg et al., 2019) or synthetic document templates (Kir- itchenko and Mohammad, 2018; Park et al., 2018). The Wikipedia Talk corpus1 (Wulczyn et al., 2017) provides de- mographic information of annotators instead of the authors, Equity Evaluation Corpus2 (Kiritchenko and Mohammad, 2018) are created by sentence templates and combinations of racial names and gender coreferences. While existing work (Davidson et al., 2019; Diaz et al., 2018) infers user demographic information (white/black, young/old) from the text, such inference is still likely to cause confounding er- The work was partially done when the first author worked as an intern at Adobe Research. 1 https://figshare.com/articles/Wikipedia_ Detox_Data/4054689 2 http://saifmohammad.com/WebPages/ Biases-SA.html rors that …

Content truncated to 500 words