UTFPR at SemEval 2020 Task 12: Identifying Offensive Tweets with Lightweight Ensembles

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 2232–2236 Barcelona, Spain (Online), December 12, 2020. 2232 UTFPR at SemEval 2020 Task 12: Identifying Offensive Tweets with Lightweight Ensembles Marcos Aur´elio Hermogenes Boriola, Gustavo Henrique Paetzold Universidade Tecnol´ogica Federal do Paran´a, Toledo-PR, Brazil marcosboriola@alunos.utfpr.edu.br, ghpaetzold@utfpr.edu.br Abstract Offensive language is a common issue on social media platforms nowadays. In an effort to address this issue, the SemEval 2020 event held the OffensEval 2020 shared task where the participants were challenged to develop systems that identify and classify offensive language in tweets. In this paper, we present a system that uses an Ensemble model stacking a BOW model and a CNN model that led us to place 29th in the ranking for English sub-task A. 1 Introduction Social media websites are widespread on the internet and they are used to facilitate the communication between people. Their usage is growing more and more (Perrin, 2015). People use these means of communication to express their opinions, to share something they like, to keep in touch with their friends and relatives, and also to get to know someone new. However, some people use social media to commit attacks against others, consequently sharing offensive content on theses platforms (Chetty and Alathur, 2018). In an effort to help identify offensive language on social media posts, we rely on Natural Language Processing (NLP) techniques, since this is an area of Computer Science and Linguists that cares about the interaction of computers and humans through natural language. And this kind of task in NLP is called Binary Text Classification, where, based on textual features, a system estimates whether or not a sentence has offensive language. There are many types of approaches to solving a Text Classification problem, such as statistical models (e.g., Naive Bayes, SVM), machine learning models (e.g., gradient boosted decision trees), deep learning models (e.g., LSTM, RNN, CNN, BERT), and other methods as well. Last year, OffensEval 2019 shared task participants created many approaches to solve the issue of offensive language identification, where 70% of them used deep learning models as a solution to this task. Analyzing only the methods used by the top-10 teams, seven of them used BERT to solve the task (Zampieri et al., 2019). Although BERT has managed to achieve state-of-the-art results in many NLP tasks (Devlin et al., 2018), it is very computationally expensive to train compared to most machine learning models (Peng, 2019). Due to its complexity and robustness, the BERT model requires many GPUs for its training and this model is not very suitable for low resource languages. In an attempt to solve this problem, we developed a more lightweight approach using an ensemble model stacking a Bag-of-words (BOW) model and a Convolutional Neural Network (CNN). The BOW model can be easily applied to even low resource languages and the CNN model is considerably less costly to train than the BERT model. We claim that’s a lightweight model due to a runtime of 11 hours in a freely available …

Content truncated to 500 words