HATEMINER at SemEval-2019 Task 5: Hate speech detection against Immigrants and Women in Twitter using a Multinomial Naive Bayes Classifier

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 404–408 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 404 HATEMINER at SemEval-2019 Task 5: Hate speech detection against Immigrants and Women in Twitter using a Multinomial Naive Bayes Classifier Nikhil Chakravartula Teradata, Hyderabad nikhil.chakravartula@gmail.com Abstract This paper describes our participation in the SemEval 2019 Task 5 - Multilingual Detection of Hate. This task aims to identify hate speech against two specific targets, immigrants and women. We compare and contrast the per- formance of different word and sentence level embeddings on the state-of-the-art classifica- tion algorithms. Our final submission is a Multinomial binarized Naive Bayes model for both the subtasks in the English version. 1 Introduction Twitter is a micro-blogging platform where peo- ple exchange ideas using short messages called tweets. Users can propagate their notions, includ- ing hatred against an individual or a group, to the entire global population with a latency of a few seconds. This poses a unique challenge of devel- oping systems that can automatically identify and mitigate hate speech. Although twitter condemns hate speech through its hateful conduct policy1, enforcing it is difficult. There are several reasons for this. Tweets often contain emoticons, emo- jis, language slangs, hashtags and other noisy data. Often, offensive and abusive language may be er- roneously perceived as hate speech and hence it is important to distinguish offensive, abusive and hateful languages (Davidson et al., 2017; Waseem and Hovy, 2016). These problems are exacerbated by the fact that even humans find it difficult to de- lineate offensive and hateful language. Many approaches have been put forward to de- tect hate speech. Bag of words and ngram features are effective in hate speech detection (Burnap and Williams, 2015; Warner and Hirschberg, 2012) as well as the detection of abusive and offensive con- tent (Nobata et al., 2016). Gitari et al. (2015) used 1https://help.twitter. com/en/rules-and-policies/ hateful-conduct-policy lexical resources to look up certain words that con- tribute significantly to hate speech but such fea- tures, when used in isolation may not be very ef- fective. SVM (Burnap and Williams, 2015), Naive Bayes (Kwok and Wang, 2013) and Logistic Re- gression (Davidson et al., 2017) are some of the classifiers used in this domain. Most of the above methods are targeted to detect general hate speech. Through this task, we aim to identify hate speech, specifically against immi- grants and women. Frenda et al. (2018) used lexi- con resources to identify misogynistic comments. Ahluwalia et al. (2018) used an ensemble of ran- dom forest, gradient boosting and logistic regres- sion with bag of words, ngram and lexical features to discern hatred against women. We did not find significant work in detection of hate speech in En- glish against immigrants. 2 Shared Task Description The SemEval 2019 Task 5 is divided into two sub- tasks. 1. Subtask A, where systems must predict whether a tweet is hateful (HS=1) against im- migrants and women. 2. Subtask B, where systems must first classify …

Content truncated to 500 words