Offensive Content Detection via Synthetic Code-Switched Text

Proceedings of the 29th International Conference on Computational Linguistics, pages 6617–6624 October 12–17, 2022. 6617 Offensive Content Detection Via Synthetic Code-Switched Text Cesa Salaam1 Franck Dernoncourt2 Trung Bui2 Danda Rawat1 Seunghyun Yoon2 1Howard University 2Adobe Research cesa.salaam@bison.howard.edu {franck.dernoncourt, bui, syoon}@adobe.com danda.rawat@howard.edu Abstract The prevalent use of offensive content in so- cial media has become an important reason for concern for online platforms (customer service chat-boxes, social media platforms, etc). Clas- sifying offensive and hate-speech content in on- line settings is an essential task in many appli- cations that needs to be addressed accordingly. However, online text from online platforms can contain code-switching, a combination of more than one language. The non-availability of labeled code-switched data for low-resourced code-switching combinations adds difficulty to this problem. To overcome this, we release a human-generated dataset containing around 10k samples for testing for three language com- binations en-fr, en-es, and en-de1 and a syn- thetic code-switched dataset containing 30k samples for training2. In this paper, we describe the process for gathering the human-generated data and our algorithm for creating synthetic code-switched offensive content data. We also introduce the results of a keyword classification baseline and a multi-lingual transformer-based classification model. 1 Introduction The use of offensive content in online settings such as chat-boxes, and social media platforms contin- ues to be a growing problem that requires address- ing. It can have negative effects on the psycho- emotional state of people (Saha et al., 2019). Of- fensive content and hate-speech continue to be a challenge to people world. As such, it is impor- tant to keep social media and other communication platforms free from offensive content. Consider- able research has been conducted on deep-learning techniques for detecting offensive language (Pitsilis et al., 2018; Mehra and Hasanuzzaman, 2020). One of the growing challenges in the field of content detection is code-switching (Aguilar et al., 2020; 1https://tinyurl.com/adobehuman 2https://tinyurl.com/adobesynthetic Qin et al., 2020; Tang et al., 2020; Chakravarthi et al., 2020). Code-switching refers to the use of two or more languages in a single conversation. Code-switching can occur inter-sententially (across sentences) and intra-sententially (within sentences). The combination of code-switching and offensive content increases the complexity of the classifi- cation task. As code-switching is a combination of multiple languages, resources for these various combinations are extremely low. This causes re- searchers to find ways to create viable synthetic data that can serve in place of real-world data for training purposes. However, the real world bench- mark test set still remains scarce. To stimulate the research, we create human- annotated testsets written in three pairs of lan- guages (en-fr, en-es, and en-de). Further, we pro- pose a method for creating a synthetic train set and show its applicability to detect human-annotated code-switched text. 2 Related Works Researchers have attempted to solve the prob- lem of synthetic data generation for various code- switching tasks. Theory Based Synthetic Code-switching Data Generation: Equivalency Theory (EC Theory) ex- plains a range of interesting code-switched patterns beyond lexical substitution. The EC Theory de- scribes a …

Content truncated to 500 words