A Dataset of Offensive German Language Tweets Annotated for Speech Acts

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 4799–4807 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 4799 A Dataset of Offensive German Language Tweets Annotated for Speech Acts Melina Plakidis1,2, Georg Rehm1,2 1 DFKI GmbH, Alt-Moabit 91c, 10559 Berlin, Germany 2 Humboldt-Universit¨at zu Berlin, Dorotheenstraße 24, 10117 Berlin, Germany Abstract We present a dataset consisting of German offensive and non-offensive tweets, annotated for speech acts. These 600 tweets are a subset of the dataset by (Struß et al., 2019) and comprises three levels of annotation, i. e., six coarse-grained speech acts, 23 fine-grained speech acts and 14 different sentence types. Furthermore, we provide an evaluation in both qualitative and quantitative terms. The dataset is made publicly available under a CC-BY-4.0 license. Keywords: Speech acts, hate speech detection, offensive language, annotation, corpus annotation 1 Introduction In recent years, research has invested a considerable amount of effort in offensive language and other phe- nomena related to online communication. Hate speech not only affects individuals or minority groups but it might also threaten social cohesion (Weber et al., 2019). Thus, the automatic detection of hate speech and offensive language1 continues to be an important and very relevant research topic. While there is a large body of research in this area, ap- proaches often merely classify text using binary labels (Burnap and Williams, 2016; Risch et al., 2021). Hate speech classification is in itself a complex task because it is highly subjective what constitutes hate speech. In addition, it is even more difficult to detect hate speech if it is expressed implicitly rather than explicitly (Palmer et al., 2020; Struß et al., 2019). Surprisingly little re- search exists on the pragmatic characteristics of offen- sive language. Pragmatics is “the study of how utter- ances have meanings in situations” (Leech, 1983, p. x). To address this gap and contribute to the improvement of hate speech detection, we analysed the pragmatic characteristics of offensive language. We conducted a speech act analysis of a dataset of German tweets that contain offensive language. We use a subset of the 2019 GermEval Shared Task on the Identification of Offensive Language dataset (Struß et al., 2019). Similar studies exist in which speech act theory (Austin, 1962; Searle, 1969) is applied to data (Juraf- sky, 1997; Zhang et al., 2011; Compagno et al., 2018; Vosoughi and Roy, 2016; Weisser, 2018). However, many of these concentrate on spoken language, often confined to restricted discourse scenarios such as the SPAADIA Trainline Corpus (Leech and Weisser, 2013) which consists of phone conversations between call- centre agents and customers concerning bookings of 1The terms “hate speech” and “offensive language” are used synonymously in this paper. It should be noted, how- ever, that these terms are not always used synonymously in the literature and that there are various differing definitions. Poletto et al. (2020) provide a corresponding overview. train tickets. Other approaches apply speech act the- ory to data from Twitter (Zhang et al., 2011; …

Content truncated to 500 words