Demographic Word Embeddings for Racism Detection on Twitter

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 926–936, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Demographic Word Embeddings for Racism Detection on Twitter Mohammed Hasanuzzaman ADAPT Centre Dublin City University Ireland Ga¨el Dias GREYC UMR 6072 Universit´e de Caen Normandie France Andy Way ADAPT Centre Dublin City University Ireland Abstract Most social media platforms grant users freedom of speech by allowing them to freely express their thoughts, beliefs, and opinions. Although this represents in- credible and unique communication op- portunities, it also presents important chal- lenges. Online racism is such an exam- ple. In this study, we present a super- vised learning strategy to detect racist lan- guage on Twitter based on word embed- ding that incorporate demographic (Age, Gender, and Location) information. Our methodology achieves reasonable classi- fication accuracy over a gold standard dataset (F1=76.3%) and significantly im- proves over the classification performance of demographic-agnostic models. 1 Introduction The advent of microblogging services has im- pacted the way people think, communicate, be- have, learn, and conduct their daily activities. In particular, the lack of regulation has made social media an attractive tool for people to express on- line their thoughts, beliefs, emotions and opinions. However, this transformative potential goes with the challenge of maintaining a complex balance between freedom of expression and the defense of human dignity (Silva et al., 2016). Indeed, some users misuse the medium to promote offensive and hateful language, which mars the experience of regular users, affects the business of online com- panies, and may even have severe real-life conse- quences (Djuric et al., 2015). In the latter case, (Priest et al., 2013; Tynes et al., 2008; Paradies, 2006b; Darity Jr., 2003) evidenced strong corre- lations between experiences of racial discrimina- tion and negative mental health outcomes such as depression, anxiety, and emotional stress as well as negative physical health outcomes such as high blood pressure and low infant birth weight. As the information contained in social media often reflects the real-world experiences of their users, there is an increased expectation that the norms of society will also apply in social media settings. As such, there is an increasing demand for social media platforms to empower users with the tools to report offensive and hateful content (Oboler and Connelly, 2014). Hateful content can be defined as ”speech or ex- pression which is capable of instilling or inciting hatred of, or prejudice towards, a person or group of people on a specified ground, including race, nationality, ethnicity, country of origin, ethno- religious identity, religion, sexuality, gender iden- tity or gender” (Gelber and Stone, 2007). While there are many forms of hate speech, racism is the most general and prevalent form of hate speech in Twitter (Silva et al., 2016). Racist speech relates to a socially constructed idea about differences be- tween social groups based on phenotype, ancestry, culture or religion (Paradies, 2006a) and covers the categories of race (e.g. black people), ethnic- ity (e.g. Chinese people), and …

Content truncated to 500 words