JCT at SemEval-2020 Task 12: Offensive Language Detection in Tweets Using Preprocessing Methods, Character and Word N-grams

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 2017–2022 Barcelona, Spain (Online), December 12, 2020. 2017 JCT at SemEval-2020 Task 12: Offensive Language Detection in Tweets using Preprocessing Methods, Character and Word N-grams Moshe Uzan1 Yaakov HaCohen-Kerner2 1Computer Science Department, Bar Ilan University, Ramat-Gan 5290002, Israel 2Jerusalem College of Technology (Lev Academic Center), Jerusalem 9116001, Israel uzanmacho@gmail.com, kerner@jct.ac.il Abstract In this paper, we describe our submissions to SemEval-2020 contest. We tackled subtask 12 - “Multilingual Offensive Language Identification in Social Media”. We developed different models for four languages: Arabic, Danish, Greek, and Turkish. We applied three supervised machine learning methods using various combinations of character and word n-gram features. In addition, we applied various combinations of basic preprocessing methods. Our best submission was a model we built for offensive language identification in Danish using Random Forest. This model was ranked at the 6th position out of 39 submissions. Our result is lower by only 0.0025 than the result of the team that won the 4th place using entirely non-neural methods. Our experiments indicate that char ngram features are more helpful than word ngram features. This phenomenon probably occurs because tweets are more characterized by characters than by words, tweets are short, and contain various special sequences of characters, e.g., hashtags, shortcuts, slang words, and typos. 1 Introduction Nowadays, social networks occupy an increasingly important place in our life. However, their rapid development has also allowed the development of increasingly offensive, hateful content that takes advantage of the anonymous nature of the cyber world. This bullying, trolling, and harassment content has become a growing concern for governments, organizations, and individuals. The use of machine learning (ML) and natural language processing (NLP) solutions to find this offensive content has been surprisingly useful. Still, the detection of offensive language from social media is not an easy research problem due to the different levels of ambiguities present in natural language and the noisy nature of social media language. In addition, social media subscribers come from linguistically diverse communities. Up to now, most research has focused on the detection of offensive language in English but works have also been written for other languages like Struß et al. (2019) for German and Basile et al. (2019) for Spanish. Task 12-A of SEMEVAL 2020 deals with the detection of offensive language in tweets in five languages including English. In this paper, we describe our models submitted to subtask 12-A for tweets written in four languages Arabic, Danish, Greek, and Turkish. The full description of task 12 is given in Zampieri et al. (2020). The structure of the rest of the paper is as follows. Section 2 introduces a background concerning offensive language, tweet classification, and data preprocessing. Section 3 describes subtask 12 and its datasets. In Section 4, we present the submitted models and their experimental results. Section 5 summarizes and suggests ideas for future research. 2 Background 2.1 Offensive language In recent years, more and more workshops on offensive content, hate speech, and aggression …

Content truncated to 500 words