TAC at SemEval-2020 Task 12: Ensembling Approach for Multilingual Offensive Language Identification in Social Media

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 2177–2182 Barcelona, Spain (Online), December 12, 2020. 2177 TAC at SemEval-2020 Task 12: Ensembling Approach for Multilingual Offensive Language Identification in Social Media Talha Anwar Mirza Omer Beg Department of Computer Science National University of Computer and Emerging Sciences Islamabad, Pakistan. {i181509,omer.beg}@nu.edu.pk Abstract Usage of offensive language on social media is getting more common these days, and there is a need of a mechanism to detect it and control it. This paper deals with offensive language detection in five different languages; English, Arabic, Danish, Greek and Turkish. We presented an almost similar ensemble pipeline comprised of machine learning and deep learning models for all five languages. Three machine learning and four deep learning models were used in the ensemble. In the OffensEval-2020 competition our model achieved F1-score of 0.85, 0.74, 0.68, 0.81, and 0.9 for Arabic, Turkish, Danish, Greek and English language tasks respectively. 1 Introduction With the growing and easy access to social media without check and balance, ethical and moral values are somehow put aside. One of the main issues is abusive behaviour, cyberbullying and offensive language. Recently a lot of work (Kumar et al., 2018) (Zampieri et al., 2019),(i Orts, 2019) was performed to tackle this problem. Most of the research work was done for languages like English and Arabic (Al-Hassan and Al-Dossari, 2019), and less work for languages like Danish, Turkish and Greek. The SemEval 2020 task 12 (Zampieri et al., 2020) consisted of identification of offensive words in different languages such as Arabic, Turkish, English, Greek and Danish. This paper proposed an ensemble approach for offensive language identification in social media text. We have participated in all languages tasks. For Arabic, Turkish, Danish and Greek language, we used FastText embeddings. For English subtask A, we used FastText embeddings combined with GloVe embeddings. The datasets were skewed, for which we used the focal loss as it gives more weightage to less occurred class in the dataset. This paper is organized as follows. Section 2 presents related work in the area of offensive language detection. Section 3 deals with more detail of provided dataset and the methodology used for the tasks. Results are discussed in section 4. Section 5 concludes the paper and presents some ideas for future work. 2 Related Work AA Altowayan and Ashraf (2017) trained sentiment specific embedding using FastText both for the bag of words (CBOW) and skip-gram models to study syntactic and semantic information of Arabic language. They used four different datasets. Embeddings were trained on one dataset and tested on the other three datasets. F1 score of 82.47%, 82.22% and 70.62% was achieved on the rest of three datasets respectively. Similarly, using skip-gram model, F1 score of 82.31%,86.25% and 72.37% was achieved. Mohaouchane et al. (2019) achieved F1 score of 83.46% by training convolution neural network on the dataset comprised of 15,000 YouTube comments in Arabic labeled as offensive or not. They used Tree- Structured Parzen Estimator (TPE) algorithm; a Bayesian …

Content truncated to 500 words