Predicting the Type and Target of Offensive Posts in Social Media

Proceedings of NAACL-HLT 2019, pages 1415–1420 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1415 Predicting the Type and Target of Offensive Posts in Social Media Marcos Zampieri1, Shervin Malmasi2, Preslav Nakov3, Sara Rosenthal4 Noura Farra5, Ritesh Kumar6 1University of Wolverhampton, UK, 2Harvard Medical School, USA 3Qatar Computing Research Institute, HBKU, Qatar, 4IBM Research, USA 5Columbia University, USA, 6Bhim Rao Ambedkar University, India m.zampieri@wlv.ac.uk Abstract As offensive content has become pervasive in social media, there has been much research in identifying potentially offensive messages. However, previous work on this topic did not consider the problem as a whole, but rather fo- cused on detecting very specific types of offen- sive content, e.g., hate speech, cyberbulling, or cyber-aggression. In contrast, here we tar- get several different kinds of offensive content. In particular, we model the task hierarchically, identifying the type and the target of offensive messages in social media. For this purpose, we complied the Offensive Language Identi- fication Dataset (OLID), a new dataset with tweets annotated for offensive content using a fine-grained three-layer annotation scheme, which we make publicly available. We discuss the main similarities and differences between OLID and pre-existing datasets for hate speech identification, aggression detection, and simi- lar tasks. We further experiment with and we compare the performance of different machine learning models on OLID. 1 Introduction Offensive content has become pervasive in social media and thus a serious concern for government organizations, online communities, and social me- dia platforms. One of the most common strategies to tackle the problem is to train systems capable of recognizing offensive content, which can then be deleted or set aside for human moderation. In the last few years, there have been several stud- ies on the application of computational methods to deal with this problem. Prior work has stud- ied offensive language in Twitter (Xu et al., 2012; Burnap and Williams, 2015; Davidson et al., 2017; Wiegand et al., 2018), Wikipedia comments,1 and Facebook posts (Kumar et al., 2018). 1http://bit.ly/2FhLMVz Previous studies have looked into different as- pects of offensive language such as the use of abusive language (Nobata et al., 2016; Mubarak et al., 2017), (cyber-)aggression (Kumar et al., 2018), (cyber-)bullying (Xu et al., 2012; Dadvar et al., 2013), toxic comments1, hate speech (Kwok and Wang, 2013; Djuric et al., 2015; Burnap and Williams, 2015; Davidson et al., 2017; Malmasi and Zampieri, 2017, 2018), and offensive lan- guage (Wiegand et al., 2018). Recently, Waseem et al. (2017) analyzed the similarities between different approaches pro- posed in previous work and argued that there was a need for a typology that differentiates between whether the (abusive) language is directed towards a specific individual or entity, or towards a gener- alized group, and whether the abusive content is explicit or implicit. Wiegand et al. (2018) further applied this idea to German tweets. They experi- mented with a task on detecting offensive vs. non- offensive tweets, and also with a second task on further sub-classifying the offensive tweets …

Content truncated to 500 words