Recognizing Explicit and Implicit Hate Speech Using a Weakly Supervised Two-path Bootstrapping Approach

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 774–782, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Recognizing Explicit and Implicit Hate Speech Using a Weakly Supervised Two-path Bootstrapping Approach Lei Gao Texas A&M University sjtuprog@tamu.edu Alexis Kuppersmith Stanford University Lex54@stanford.edu Ruihong Huang Texas A&M University huangrh@cse.tamu.edu Abstract In the wake of a polarizing election, social media is laden with hateful content. To address various limitations of supervised hate speech classification methods includ- ing corpus bias and huge cost of annota- tion, we propose a weakly supervised two- path bootstrapping approach for an online hate speech detection model leveraging large-scale unlabeled data. This system significantly outperforms hate speech de- tection systems that are trained in a super- vised manner using manually annotated data. Applying this model on a large quan- tity of tweets collected before, after, and on election day reveals motivations and patterns of inflammatory language. 1 Introduction Following a turbulent election season, 2016’s dig- ital footprint is awash with hate speech. Apart from censorship, the goals of enabling computers to understand inflammatory language are many. Sensing increased proliferation of hate speech can elucidate public opinion surrounding polarizing events. Identifying hateful declarations can bolster security in revealing individuals harboring mali- cious intentions towards specific groups. Recent studies on supervised methods for on- line hate speech detection (Waseem and Hovy, 2016; Nobata et al., 2016) have relied on manu- ally annotated data sets, which are not only costly to create but also likely to be insufficient to ob- tain wide-coverage hate speech detection systems. This is mainly because online hate speech is rela- tively infrequent (among large amounts of online contents) and tends to transform rapidly following a new trigger event. Our pilot annotation experi- ment with 5,000 randomly selected tweets shows that around 0.6% (31 tweets) of tweets are hateful. The mass-scale (Yahoo!Finance online comments) hate speech annotation effort from Yahoo! (No- bata et al., 2016) revealed that only 5.9% of on- line comments contained hate speech. Therefore, large amounts of online texts need to be annotated to adequately identify hate speech. In recent stud- ies (Waseem and Hovy, 2016; Kwok and Wang, 2013), the data selection methods and annotations are often biased towards a specific type of hate speech or hate speech generated in certain scenar- ios in order to increase the ratio of hate speech content in the annotated data sets, which how- ever made the resulting annotations too distorted to reflect the true distribution of hate speech. Fur- thermore, inflammatory language changes dramat- ically following new hate “trigger” events, which will significantly devalue annotated data. To address the various limitations of super- vised hate speech detection methods, we present a weakly supervised two-path bootstrapping ap- proach for online hate speech detection that re- quires minimal human supervision and can be eas- ily retrained and adapted to capture new types of inflammatory language. Our two-path bootstrap- ping architecture consists of two learning compo- nents, an explicit slur term learner and …

Content truncated to 500 words