YNU NLP at SemEval-2019 Task 5: Attention and Capsule Ensemble for Identifying Hate Speech

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 529–534 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 529 YNU NLP at SemEval-2019 Task 5: Attention and Capsule Ensemble for Identifying Hate Speech Bin Wang, Haiyan Ding∗ School of Information Science and Engineering Yunnan University, Yunnan, P.R. China hyding@ynu.edu.cn Abstract This paper describes the system submitted to SemEval 2019 Task 5: Multilingual detection of hate speech against immigrants and wom- en in Twitter (hatEval). Its main purpose is to conduct hate speech detection on Twitter, which mainly includes two specific different targets, immigrants and women. We partici- pate in both subtask A and subtask B for En- glish. In order to address this task, we devel- ope an ensemble of an attention-LSTM model based on HAN and a BiGRU-capsule model. Both models use fastText pre-trained embed- dings, and we use this model in both subtasks. In comparison to other participating teams, our system is ranked 16th in the Subtask A for En- glish, and 12th in the Subtask B for English. 1 Introduction In recent years, the popularity of social network- ing and microblogging sites has increased, attract- ing more and more users. With this huge user base, social media will continue to release a large num- ber of user-generated content. As the use of social media has grown, other undesirable phenomena and behaviors have emerged. Social media users often abuse this freedom to spread abuse or hate- ful posts or comments. In many cases, these user- generated content is inherently offensive or proac- tive, and users may have to deal with threats such as cyber attacks or cyberbullying, as well as oth- er undesirable phenomena (Whittaker and Kowal- ski, 2015). So the problem of detecting and possi- bly limiting the spread of hate speech is becoming more and more important. In order to solve the problem of abuse of lan- guage in social media platforms, some related research has been published, such as cyberbul- lying (Dadvar et al., 2013), hate speech (Warn- er and Hirschberg, 2012) and abusive language ∗Corresponding author (Chen et al., 2012), most methods are based on surveillance methods (Schmidt and Wiegand, 2017). There are also some (racial discrimination) bias towards specific goals. In (Waseem and Hov- y, 2016), the authors proposed a series of criteria based on critical race theory to identify racism and gender discrimination, they use n-gram models for research; Tulkens et al. studied racism detection in Dutch social media (Tulkens et al., 2016). A recent discussion of the challenge of identifying hate speech was proposed by Kumar et al. (Kumar et al., 2018). The results show that it is difficult to distinguish between open and covert attacks in social media. SemEval 2019 Task 5 is proposed to identify hate speech about immigrants and women in Twit- ter for English or Spanish, and classify hate speech and judge whether the target is an individual or a group (Basile et al., 2019). Hate speech is of- ten defined as …

Content truncated to 500 words