LT3 at SemEval-2019 Task 5: Multilingual Detection of Hate Speech Against Immigrants and Women in Twitter (hatEval)

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 436–440 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 436 LT3 at SemEval-2019 Task 5: Multilingual Detection of Hate Speech Against Immigrants and Women in Twitter (hatEval) Nina Bauwelinck, Gilles Jacobs, V´eronique Hoste and Els Lefever LT3, Language and Translation Technology Team Department of Translation, Interpreting and Communication – Ghent University Groot-Brittanni¨elaan 45, 9000 Ghent, Belgium firstname.lastname@ugent.be, gillesm.jacobs@ugent.be Abstract This paper describes our contribution to the SemEval-2019 Task 5 on the detection of hate speech against immigrants and women in Twitter (hatEval). We considered a super- vised classification-based approach to detect hate speech in English tweets, which combines a variety of standard lexical and syntactic fea- tures with specific features for capturing offen- sive language. Our experimental results show good classification performance on the train- ing data, but a considerable drop in recall on the held-out test set. 1 Introduction The exponential growth of social media such as Twitter, Facebook, Youtube and community fo- rums has created a variety of novel ways for all of us to communicate with each other, but this op- portunity to freely communicate online has unfor- tunately also given a forum to people who want to denigrate others because of their race, colour, gen- der, sexual orientation, religion, etc. While there has been an increasing interest in automatic hate speech detection in social media, the problem is far from solved, partly due to the low consensus on what exactly constitutes hate speech, how it relates to offensive language and bullying and thus the low reliability of hate speech annotations (Ross et al., 2017). Davidson et al. (2017) for exam- ple observe that their classifications of hate speech tend to reflect their own subjective biases: while racist and homophobic insults are considered hate- ful, they tend to see sexist language as merely offensive. When we consider the different ap- proaches that address hate speech, we can observe that -apart from simple methodologies that rely on lookup in a dictionary of hateful terms (Tulkens et al., 2016) - most methods cast the problem as a supervised classification task either using a more standard machine learning approach or deep learn- ing methods (Pitsilis et al., 2018). This was also the approach we took for our hate speech detec- tion system. We participated for both subtasks proposed for English for Task 5 (Basile et al., 2019), being TASK A, which was defined as a binary classifica- tion task where systems have to predict whether a tweet with a given target (women or immigrants) is hateful or not hateful, and TASK B, where systems are asked first to classify hateful tweets as aggres- sive or not aggressive, and second to identify the target harassed as individual or generic (i.e. single human or group). 2 System Description We designed a cascaded classification-based ap- proach, where a first classifier categorizes a tweet as being hateful or not, while in a second step the hateful tweets are classified as …

Content truncated to 500 words