Emoji Powered Capsule Network to Detect Type and Target of Offensive Posts in Social Media

Proceedings of Recent Advances in Natural Language Processing, pages 474–480, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_056 474 Emoji Powered Capsule Network to Detect Type and Target of Offensive Posts in Social Media Hansi Hettiarachchi1, Tharindu Ranasinghe2 1Department of Computer Science and Engineering,, University of Moratuwa, Sri Lanka 2Research Group in Computational Linguistics, University of Wolverhampton, UK hansi.11@cse.mrt.ac.lk t.d.ranasinghehettiarachchige@wlv.ac.uk Abstract This paper describes a novel research ap- proach to detect type and target of offen- sive posts in social media using a cap- sule network. The input to the network was character embeddings combined with emoji embeddings. The approach was evaluated on all the subtasks in SemEval- 2019 Task 6: OffensEval: Identifying and Categorizing Offensive Language in So- cial Media. The evaluation also showed that even though the capsule networks have not been used commonly in NLP tasks, they can outperform existing state of the art solutions for offensive language de- tection in social media. 1 Introduction Social media has become a normal medium of communication for people these days as it pro- vides the convenience of sending messages fast from a variety of devices. Unfortunately, social networks also provide the means for distributing abusive and aggressive content. Given the amount of information generated every day on social me- dia, it is not possible for humans to identify and remove such messages manually, instead it is nec- essary to employ automatic methods. Recently, many shared tasks have been introduced to encour- age the development of methods capable of clas- sifying messages from social media as offensive. As an example, the First Workshop on Trolling, Aggression and Cyberbullying has organised the First Shared Task on Aggression Identification to classify messages from Facebook and Twitter into three categories Overtly Aggressive (OAG), Covertly Aggressive (CAG) and Non-aggressive (NAG) (Kumar et al., 2018). The task was organ- ised for English and Hindi. Recently, more complete dataset covering dif- ferent aspects of offensive identification was re- leased for the shared task in SemEval-2019 Task 6: OffensEval: Identifying and Categorizing Of- fensive Language in Social Media. The task was not only to identify offensive messages in social media. The participants had to categorize the of- fensive language and also had to identify the tar- geted audience (Zampieri et al., 2019). We used this dataset to experiment our novel architecture since it covers more aspects in offensive language detection in social media. More details about the 2.tasks and the dataset will be discussed in Section People from all over the world uses social me- dia. Therefore, a random sample of social media messages can be written in several languages. As a result, systems that detect offensive posts with- out relying too much on language dependent fea- tures would be valuable in real life scenarios. In the offensive language identification shared tasks too, most researches have worked on systems that rely on word/character embeddings rather than lin- guistics features. As an example Galery and Char- itos (2018) has taken an approach to feed fast- text (Mikolov et al., 2018) …

Content truncated to 500 words