UA at SemEval-2019 Task 5: Setting A Strong Linear Baseline for Hate Speech Detection

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 508–513 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 508 UA at SemEval-2019 Task 5: Setting a Strong Linear Baseline for Hate Speech Detection Carlos Perell´o♦ David Tom´as♦ Alberto Garcia-Garcia♦ Jose Garcia-Rodriguez♦ Jose Camacho-Collados♣ ♦ University of Alicante, Spain ♣ Cardiff University, United Kingdom ♦cpc69@alu.ua.es,dtomas@dlsi.ua.es,{agarcia,jgarcia}@dtic.ua.es, ♣camachocolladosj@cardiff.ac.uk Abstract This paper describes the system developed at the University of Alicante (UA) for the Se- mEval 2019 Task 5: Multilingual detection of hate speech against immigrants and women in Twitter. The purpose of this work is to build a strong baseline for hate speech detec- tion by means of a traditional machine learn- ing approach with standard textual features, which could serve as a reference to compare with deep learning systems. We participated in both task A (Hate Speech Detection against Immigrants and Women) and task B (Aggres- sive behavior and Target Classification) for both English and Spanish. Given the text of a tweet, task A consists of detecting hate speech against women or immigrants in the text, whereas task B consists of identifying the target harassed as individual or generic, and to classify hateful tweets as aggressive or not aggressive. Despite its simplicity, our system obtained a remarkable macro-F1 score of 72.5 (sixth highest) and an accuracy of 73.6 (sec- ond highest) in Spanish (task A), outperform- ing more complex neural models from a total of 40 participant systems. 1 Introduction Due to the massive rise of users in social me- dia, the presence of verbal abuse, hate speech and bully-attitudes has increased over the years. clear example is Twitter, where users find waysA to anonymously harass and offend other individ- uals or collectives. This is especially troublesome as hate speech and hate crime are strongly related. Therefore, an early detection of hate speech could help prevent the subsequent hate crime. Online platforms like Twitter have been seeking to com- bat hate speech on their site, but it still requires a lot of manual work because there is not a reli- able automatic method that can correctly identify hate speech behaviour. Building such automatic (or semi-automatic) systems is therefore essential to effectively fight this problem. Hate speech detection is still a challenging task due to a number of reasons. First, hate speech con- tent tends to be ambiguous and context-dependant (Chatzakou et al., 2017). Moreover, hate speech can cross sentence boundaries and be present in sarcastic comments in the same voice as the peo- ple that were producing abusive languages. These and other issues for detecting hate speech are dis- cussed in more detail in Nobata et al. (2016). In order to deal with these issues, over the past few years several techniques to detect hate speech and abusive language online have been proposed.1 Previous works made use of heterogeneous fea- tures such as bag of words, n-grams, punctuation, as well as lexical features and user-related fea- tures (Chatzakou et al., 2017). In addition to these features, previous …

Content truncated to 500 words