SalamNET at SemEval-2020 Task 12: Deep Learning Approach for Arabic Offensive Language Detection

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 2133–2139 Barcelona, Spain (Online), December 12, 2020. 2133 SalamNET at SemEval-2020 Task 12: Deep Learning Approach for Arabic Offensive Language Detection Fatemah Husain Kuwait University / State of Kuwait f.husain@ku.edu.kw Jooyeon Lee∗ George Mason University / USA jlee252@masonlive.gmu.edu Sam Henry George Mason University / USA shenry20@masonlive.gmu.edu Özlem Uzuner George Mason University / USA ouzuner@masonlive.gmu.edu Abstract This paper describes SalamNET, an Arabic offensive language detection system that has been submitted to SemEval 2020 shared task 12: Multilingual Offensive Language Identification in Social Media. Our approach focuses on applying multiple deep learning models and conducting in depth error analysis of results to provide system implications for future development consider- ations. To pursue our goal, a Recurrent Neural Network (RNN), a Gated Recurrent Unit (GRU), and Long-Short Term Memory (LSTM) models with different design architectures have been de- veloped and evaluated. The SalamNET, a Bi-directional Gated Recurrent Unit (Bi-GRU) based model, reports a macro-F1 score of 0.83. 1 Introduction Online offensive language is a critical problem that is threatening the safety and well-being of society. Previous studies demonstrate the connection between different forms of online offensive language, such as hate speech, cyberbullying, and real-world violence (Sap et al., 2019). The attack on the Tree of Life synagogue on October 2018 in Pittsburgh, which was described as the deadliest attack in the Jewish community in the US, shows the severe effects of online hate speech as the shooter had an online profile full of hate speech language (Keyser, 2018). Given the significance of the impact online offensive language can have, further investigation of the problem is warranted. We hypothesize that developing a Natural Language Processing (NLP) approach with language specific features can help in detecting offensive language to stop these attacks. While there are several studies on English offensive language detection, very few studies are available for the Arabic language. The Arabic language has specific language characteristics and features that require distinct system development (Mubarak et al., 2020). The main goal of this paper is to develop a system for online Arabic offensive language detection. We participated in SemEval 2020 shared task 12: Multilingual Offensive Language Identification in Social Media with two submissions, a Logistic Regression (LR) based model and a Recurrent Neural Net- work (RNN) based model. We evaluate the performance of both models through a detailed error analysis to enhance the performance until we arrive to the SalamNET system for online Arabic offensive language detection from Twitter. We explore multiple deep learning models, including an RNN, a Gated Recurrent Unit (GRU), a Bi-directional Gated Recurrent Unit (Bi-GRU), Long-Short Term Memory (LSTM), and a Bi-directional Long-Short Term Memory (Bi-LSTM) models separately. Based on our findings, the ∗Second first author This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/. 2134 highest macro-F1 score among all models is 0.83, which was reported by the SalamNET that consists of a Bi-GRU based model with Term Frequency-Inverse Document Frequency (TF-IDF) …

Content truncated to 500 words