YNUWB at SemEval-2019 Task 6: K-max pooling CNN with average meta-embedding for identifying offensive language

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 818–822 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 818 YNUWB at SemEval-2019 Task 6: K-max pooling CNN with average meta-embedding for identifying offensive language Bin Wang, Xiaobing Zhou∗ ,Xuejie Zhang School of Information Science and Engineering Yunnan University, Yunnan, P.R. China ∗Corresponding authorzhouxb@ynu.edu.cn Abstract This paper describes the system submitted to SemEval 2019 Task 6: OffensEval 2019. The task aims to identify and categorize offen- sive language in social media, we only par- ticipate in Sub-task A, which aims to iden- tify offensive language. In order to address this task, we propose a system based on a K-max pooling convolutional neural network model, and use an argument for averaging as a valid meta-embedding technique to get a meta- embedding. Finally, we use a cyclic learn- ing rate policy to improve model performance. Our model achieves a Macro F1-score of 0.802 (ranked 9/103) in the Sub-task A. 1 Introduction In the past ten years, with the popularity of the Internet, social media platforms such as facebook and twitter have gradually become important tools for people’s daily communication, and users can publish their own content on these platforms. As the number of people interacting on social medi- a platforms increases, online aggression language behavior also grows, and now it has become a ma- jor source of social conflict. Semeval 2019 Task 6 is proposed for identify- ing online offensive languages (Zampieri et al., 2019b). Its goal is to use computational method- s to identify offense, aggression and hate speech in user-generated content on online social medi- a platforms. We can prevent abuse of offensive language by using this approach in social media platforms. This task gives us some data from the social media platform, and classifies the content through computational analysis. In this competition, we only participate in Sub- task A: identification of offensive language. For this task, we use a deep learning method to build a K-max pooling convolutional neural network model which uses convolutional neural networks of different filters to extract features and preserves k largest eigenvalues during the pooling phase. We use two different pre-training vector model- s (fastText and Glove) to obtain a more accurate meta-embedding with a simple averaging tech- nique (Coates and Bollegala, 2018). In addition, we have adopted a Cyclic Learning Rate (CLR) strategy (Smith, 2017), which avoids the process to find the optimal learning rate, and the learning rate varies within a reasonable interval rather than monotonically. Unlike the Adative Learning Rate, the CLR does not require additional calculations. The rest of the paper is structured as follows: In section 2, we describe some of the relevant research work. In section 3, we describe the task data and how to build the model. In section 4, we describe the experimental results. 2 Related Work In recent years, offensive language has prevailed in social media, and people are increasingly inter- ested in identifying offensive speech, especially on social media platforms. …

Content truncated to 500 words