INGEOTEC at SemEval-2020 Task 12: Multilingual Classification of Offensive Text

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1992–1997 Barcelona, Spain (Online), December 12, 2020. 1992 INGEOTEC at SemEval-2020 Task 12: Multilingual Classification of Offensive Text Sabino Miranda-Jim´enez and Eric S. Tellez and Mario Graff CONACyT - INFOTEC, Aguascalientes, M´exico {mario.graff,sabino.miranda,eric.tellez}@infotec.mx Daniela Moctezuma CONACyT - CentroGEO, Aguascalientes, M´exico dmoctezuma@centrogeo.edu.mx Abstract This paper describes our participation in OffensEval challenges for English, Arabic, Danish, Turkish, and Greek languages. We used several approaches, such as µTC, TextCategorization, and EvoMSA. Best results were achieved with EvoMSA, which is a multilingual and domain- independent architecture that combines the prediction from different knowledge sources to solve text classification problems. 1 Introduction Freedom of expression is one of the important factors that make social media platforms popular; there, people interact with others and express themselves freely. However, offensive content has become per- vasive in social media that could mock or insult both individuals or groups of people. Thus, detecting offenses and misbehavior expressed in text form is useful to measure the people’s feelings and warn them about possible attacks on others such as abusive language, hate speech, cyberbullying, trolling, among others social problems (Waseem et al., 2017; Zampieri et al., 2019a). The scientific community organize several challenges periodically, like OffensEval, to tackle these text classification problems. In 2019, OffensEval (Zampieri et al., 2019c) organized three tasks for English, and this year (OffensEval-2020) (Zampieri et al., 2020) are included Arabic, Greek, Danish, and Turkish languages. The tasks consist in identifying the type of offensive language in short texts. In this paper, we present the results of our participation in all tasks and all languages previously mentioned. OffensEval challenge consists in determining whether a given message has offensive content. It is divided into three tasks A, B, and C for English and only task A for Arabic, Greek, Danish, and Turkish. Task A is dedicated to identifying the offensive language, i.e., determine if a message is offensive or not offensive. Task B is about categorizing offense types; that is, a tweet containing an insult or threat to someone, or a tweet containing non-targeted profanity and swearing. Finally, task C focuses on identifying the target, i.e., whether the offensive content is about an individual, a group, or others. Abusive and offensive language identification problems includes aggression (Kumar et al., 2018; Arag´on et al., 2019), hate speech (Basile et al., 2019), cyberbullying (Smith et al., 2008; Arroyo- Fern´andez et al., 2018), and offenses on targets (individuals or groups) (Waseem et al., 2017; Zampieri et al., 2019c). Typically, these problems are tackled as supervised learning problems, i.e., classification. For example, to identify the offensive language, the work by Liu et al. (2019) uses pre-trained BERT with fine-tuning on the training dataset. The approach described by Nikolov and Radivchev (2019) uses pre-trained BERT and GloVe vectors, along with techniques for overcoming unbalanced class distribu- tion in the provided test data, authors show how these techniques increase the performance in general. Kebriaei et al. (2019) study the performance of combining TF-IDF weighting, …

Content truncated to 500 words