“Why do I feel offended?” - Korean Dataset for Offensive Language Identification

Findings of the Association for Computational Linguistics: EACL 2023, pages 1142–1153 May 2-6, 2023 ©2023 Association for Computational Linguistics “Why do I feel offended?’ Korean Dataset for Offensive Language Identification San-Hee Park1∗ Kang-Min Kim2∗ O-Joun Lee 2 Youjin Kang1 Jaewon Lee3 Su-Min Lee2 SangKeun Lee1 1 Korea University, Seoul, Republic of Korea 2 The Catholic University of Korea, Bucheon, Republic of Korea 3 Seoul National University, Seoul, Republic of Korea carpediem20@korea.ac.kr, {kangmin89, ojlee}@catholic.ac.kr yjkang10@korea.ac.kr enotchi@snu.ac.kr, sumini0516@catholic.ac.kr, yalphy@korea.ac.kr Abstract Warning: This paper contains some offensive expressions. Offensive content is an unavoidable issue on social media. Most existing offensive language identification methods rely on the compilation of labeled datasets. However, existing meth- ods rarely consider low-resource languages that have relatively less data available for training (e.g., Korean). To address these issues, we con- struct a novel KOrean Dataset for Offensive Language Identification (KODOLI). KODOLI comprises more fine-grained offensiveness cat- egories (i.e., not offensive, likely offensive, and offensive) than existing ones. A likely offens- ive language refers to texts with implicit of- fensiveness or abusive language without offens- ive intentions. In addition, we propose two auxiliary tasks to help identify offensive lan- guages: abusive language detection and senti- ment analysis. We provide experimental res- ults for baselines on KODOLI and observe that pre-trained language models suffer from identi- fying "LIKELY" offensive statements. Quant- itative results and qualitative analysis demon- strate that jointly learning offensive language, abusive language and sentiment information im- proves the performance of offensive language identification. 1 Introduction Data-driven approaches for detecting and measur- ing offensive content have steadily grown from stat- istical methodologies to deep learning models for natural language processing (Balayn et al., 2021). Although various methods for detecting offensive language have been proposed, most of them rely on composing training datasets to determine whether a statement is offensive (Fortuna and Nunes, 2018; Mishra et al., 2019; Vidgen and Derczynski, 2020). In South Korea, most of the population actively ∗ These authors contributed equally to this work. Figure 1: Understanding offensive text (a) and (b) in real-world scenarios considering three questions: identi- fication of offense, existence of abusive language, and underlying sentiment with intention. We supplement the description with examples. uses the Internet, and the size of online communit- ies is large compared with the population (Park et al., 2021b). The social problems caused by of- fensive comments have also increased (BBC, 2022). Therefore, we need to analyze and discuss Korean texts and their offensiveness. Recent approaches have been studied to under- stand offensive language based on the typology of (Waseem et al., 2017), which differentiates whether the abusive language is directed to a specific indi- vidual or group, and whether it is explicit or im- plicit (Zampieri et al., 2019a; Caselli et al., 2020). This typology helps to identify the offensive lan- guage from the statement. However, most exist- ing studies (Sigurbergsson and Derczynski, 2019; Zampieri et al., 2019b) have considered the offens- 1142 ive language detection problem as a binary classi- fication task for distinguishing offensive languages. …

Content truncated to 500 words