Finding Microaggressions in the Wild: A Case for Locating Elusive Phenomena in Social Media Posts

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1664–1674, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1664 Finding Microaggressions in the Wild: A Case for Locating Elusive Phenomena in Social Media Posts Luke M. Breitfeller♠ Emily Ahn♠ David Jurgens♦ Yulia Tsvetkov♠ ♠Carnegie Mellon University ♦University of Michigan {mbreitfe, eahn1, ytsvetko}@cs.cmu.edu, jurgens@umich.edu Abstract Microaggressions are subtle, often veiled, manifestations of human biases. These un- civil interactions can have a powerful negative impact on people by marginalizing minorities and disadvantaged groups. The linguistic sub- tlety of microaggressions in communication has made it difficult for researchers to analyze their exact nature, and to quantify and extract microaggressions automatically. Specifically, the lack of a corpus of real-world microag- gressions and well-defined criteria for annotat- ing them have prevented researchers from ad- dressing these problems at scale. In this pa- per, we devise a general but nuanced, com- putationally operationalizable typology of mi- croaggressions based on a small subset of mi- croaggression data that we have. We then cre- ate two datasets: one with examples of di- verse types of microaggressions recollected by their targets, and another with gender-based microaggressions in public conversations on social media. We introduce a new, more ob- jective criterion for annotation and an active- learning based procedure that increases the likelihood of surfacing posts containing mi- croaggressions. Finally, we analyze the trends that emerge from these new datasets. 1 Introduction Toxicity and offensiveness are not always ex- pressed with toxic language. While a substantial community effort has rightfully focused on iden- tifying, preventing, and mitigating overtly toxic, profane, and hateful language (Schmidt and Wie- gand, 2017), offensiveness spans a far larger spec- trum that includes comments with more implicit and subtle signals that are no less offensive (Ju- rgens et al., 2019). One significant class of subtle-but-offensive comments includes microag- gressions (Sue et al., 2007, MAS), defined in Merriam-Webster as “a comment or action that Figure 1: Existing state-of-the-art tools for hate speech de- tection and sentiment analysis cannot identify the veiled of- fensiveness of microaggressions (MAS) like the one in this real comment, because in many cases, the framing of a MA includes stylistic markers of positive language. This moti- vates the need for a corpus and new tools for detecting MAS. subtly or often unconsciously expresses a preju- diced attitude toward a member of a marginalized group such as a racial minority.” Though sub- tle, MAS have been shown to have lasting harm- ful impacts on their targets. Qualitative interviews suggest that the subtlety of MAS may cause even greater levels of situational stress than overt ag- gression (Sue, 2010; Nadal et al., 2014). Despite a public effort to recognize and reduce—if not eliminate—their occurrence (Kim, 2013; Neff, 2015), there has been no computa- tional work to detect and analyze MAS at scale. Instead, much of the recent work has focused on explicitly toxic language (e.g., Waseem et al., …

Content truncated to 500 words