IR3218-UI at SemEval-2020 Task 12: Emoji Effects on Offensive Language IdentifiCation

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 1998–2005 Barcelona, Spain (Online), December 12, 2020. 1998 IR3218-UI at SemEval-2020 Task 12: Emoji Effects on Offensive Language Identification Sandy Kurniawan1, Indra Budi2, Muhammad Okky Ibrohim3 Faculty of Computer Science Universitas Indonesia Kampus UI, Depok, 16424, Indonesia 1sandy.kurniawan@ui.ac.id, {2indra, 3okkyibrohim}@cs.ui.ac.id Abstract In this paper, we present our approach and the results of our participation in OffensEval 2020. There are three sub-tasks in OffensEval 2020, namely offensive language identification (sub-task A), automatic categorization of offense types (sub-task B), and offense target identification (sub- task C). We participated in sub-task A of English OffensEval 2020. Our approach emphasizes on how the emoji affects offensive language identification. Our model used LSTM combined with GloVe pre-trained word vectors to identify offensive language on social media. The best model obtained macro F1-score of 0.88428. 1 Introduction Offensive language has become severe problems with the rapid growth of people using social media platforms such as Twitter. On social media, users can easily express their opinion about any topic which leads to some users posting offensive content while engaging in social media. The offensive contents sometimes are used to target another user, whether individual or particular groups based on the user’s views. An automatic method of identifying offensive language is needed to prevent the spread of offensive content on social media. Zampieri et al. (2020) organized the OffensEval 2020: Multilingual Offensive Language Identification in Social Media as Task 12 on SemEval-2020. OffensEval 2020 covers multilingual language consists of Arabic, Danish, English, Greek, and Turkish. The main focus of OffensEval 2020 is to solve problems on how to identify the offensive language (sub-task A), offense types categorization (sub-task B), and offense target identification (sub-task C). Sub-task A ran for all languages while sub-task B and sub-task C only ran for English. Our approach used a neural network model based on Long Short-Term Memory (LSTM) in order to identify offensive language on English tweets. This paper mainly describes our submission for OffensEval-2020 Sub-task A, but our model can be further developed to learn other sub-tasks. This paper is organized as follows. Related work has been discussed in section 2, dataset and metho- dology have been described in section 3. The experiment results and discussions described in section 4. Section 5 concludes the research and the future works of the research. 2 Related Work Many research on offensive language identification has been done. Wiedemann et al. (2018) studied the potential of transfer learning in automatic offensive language detection. They employ many transfer learning scenarios to train their neural network. The results proved that transfer learning improves offensive language detection performance. Ramakrishnan et al. (2019) used an ensemble model based on logistic regression and tree-based model to identify offensive language on SemEval-2019 Task 6. Char n-grams, word n-grams, part of speech and GloVe embedding were used as features. By combining 5 different models with the varying feature set as input, an ensemble model was built. Using the ensemble model, the offensive …

Content truncated to 500 words