JTML at SemEval-2019 Task 6: Offensive Tweets Identification using Convolutional Neural Networks

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 657–661 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 657 JTML at SemEval-2019 Task 6: Offensive Tweets Identification using Convolutional Neural Networks Johnny Torres ESPOL University jomatorr@espol.edu.ec Carmen Vaca ESPOL University cvaca@fiec.espol.edu.ec Abstract In this paper, we propose the use of a Convo- lutional Neural Network (CNN) to identify of- fensive tweets. We use an end-to-end model (i.e., no preprocessing) and fine-tune pre- trained embeddings (FastText) during training for learning words’ representation. We com- pare the proposed CNN model to a baseline model, such as Linear Regression, and several neural models. The results show that CNN outperforms other models, and stands as a sim- ple but strong baseline in comparison to other systems submitted to the Shared Task. 1 Introduction The fast growth of online social networks (OSNs) has provided a medium for users to express their ideas and opinions about any topic. However, some users post offensive content which may de- ter other users from engaging in online discus- sions. Despite the tools provided by some OSNs to block other users and report offensive content, the manual verification of these events are limited in scale and costs due to a large number of malicious events performed by users or bots. Therefore, it is critical to developing automated tools to mod- erate the content that are robust to ambiguity, sar- casm, and adversarial attacks (Fortuna and Nunes, 2018). Offensive language detection is an active re- search area, and several research efforts aim to contribute datasets, propose taxonomies, and im- prove current models to identify offensive con- tent. In this direction, Zampieri et al. (2019b) proposed a shared task for Identifying and Cat- egorizing Offensive Language in Social Media. The shared task is composed of the following sub- tasks: a) Offensive language identification, b) Au- tomatic categorization of offense types, and c) Of- fense target identification. tweet ASubtask B C If the tournament of shit ain’t on here... OFF UNT - He is so full of BS! OFF TIN IND swear niggas make me wanna turn this phone off OFF TIN GRP Kick the absolute shite out of the car. OFF TIN OTH Table 1: Examples of Offensive Tweets in the dataset. Table 1 shows some examples in the dataset of the shared task, and the labels in each of the sub- tasks. The labels indicate if the tweet is Offensive (OFF) and if it is an untargeted (UNT) or targeted (TIN) offense. The targets of the offensive tweets are individual (IND), group (GRP), other (OTH). This paper contributes specifically to the subtask A in the shared task. The unstructured and noisy nature of user- generated content on OSNs poses a challenge for classification models. Traditional approaches use a sparse representation for text data, such as the bag of words (BOW) or TF-IDF (Manning et al., 2008). We propose a model based on Convolutional Neural Networks (CNN) to identify and catego- rize offensive language on tweets. The …

Content truncated to 500 words