ROFF - A Romanian Twitter Dataset for Offensive Language

Proceedings of Recent Advances in Natural Language Processing, pages 895–900 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_102 895 ROFF - A Romanian Twitter Dataset for Offensive Language Mihai Manolescu University of T¨ubingen mihai.manolescu@student. uni-tuebingen.de C¸ a˘grı C¸ ¨oltekin University of T¨ubingen ccoltekin@sfs. uni-tuebingen.de Abstract This paper describes the annotation process of an offensive language data set for Roma- nian on social media. To facilitate comparable multi-lingual research on offensive language, the annotation guidelines follow some of the recent annotation efforts for other languages. The final corpus contains 5000 micro-blogging posts annotated by a large number of volunteer annotators. The inter-annotator agreement and the initial automatic discrimination results we present are in line with earlier annotation ef- forts. 1 Introduction The use of words to hurt others is a curious aspect of natural languages. Besides the scientific curios- ity, however, certain forms of offensive language can be harmful to individuals, may have discrimi- natory, toxifying effects on the society, and can be against law. An annotated corpus of offensive lan- guage is, hence, a valuable resource for understand- ing, identifying and preventing offensive content, particularly in online communication. This paper describes such an annotated corpus of Romanian on social media. The study of offensive content online goes back to the earlier days of the Internet (Lea et al., 1992; Kayany, 1998). However, identifying var- ious forms of offensive or abusive language online, such as hate speech or verbal harassment, has at- tracted considerable recent interest. The increased interest is evidenced by a number of recent shared tasks (Kumar et al., 2018; Wiegand et al., 2018; Zampieri et al., 2019b; Basile et al., 2019) to the extent that OffensEval 2020 shared task in SemeE- val evaluation campaign (Zampieri et al., 2020) received submissions from 145 teams (out of 528 registered teams). An important ingredient of these studies is an an- notated corpus. Recent annotations efforts with the aim of detecting offensive language typically focus on a particular form of offensive content, such as hate speech (Agarwal and Sureka, 2015; David- son et al., 2017; Del Vigna et al., 2017; ElSherief et al., 2018; Gao and Huang, 2017; Gitari et al., 2015; Sanguinetti et al., 2018a; Waseem, 2016, just to name a few), or cyberbullying (Xu et al., 2012; Dadvar et al., 2013; Dinakar et al., 2012; Nitta et al., 2013; Van Hee et al., 2015). However, it is often difficult to study a particular form of offensive lan- guage in isolation (Malmasi and Zampieri, 2018; Vidgen and Derczynski, 2020). Fore example, a classifier trained on a data set that annotates only in- stances of hate speech may detect other forms of of- fensive language as hate speech. As a result, some of the recent studies annotate multiple forms of offensive language together (Wiegand et al., 2018; Struß et al., 2019; Zampieri et al., 2019a). In partic- ular, the annotation scheme for the OLID data set of English tweets by Zampieri et al. (2019a) has been used for annotating a variety of languages, includ- ing Arabic …

Content truncated to 500 words