A Deep Learning Framework for Auto Detection of Hate Speech on Social Media

Abstract

Identifying hate speech on social media is important work. Unbridled development of anger can seriously harm marginalized individuals or groups, as well as our society as a whole. Social media is the main platform for spreading hate speech online. This greatly increases the complexity of automatic detection, as social media posts often use poorly written text and paralinguistic characters such as emoticons and hashtags. The context-dependent nature of the task and the lack of consensus on what constitutes hate speech add additional challenges, making it even for humans. This makes the creation of large tagged corpora complex and resource-intensive. The main problem was ungrammatical writing. Recently introduced Deep Neural Network (DNN) models, which have the ability to quickly learn various features, have greatly reduced the problem of grammatically incorrect text. To automatically identify hate speech from social media data, we developed a deep natural language processing (NLP) model that combines convolutional and recurrent layers. Using our model on the HASOC2019 corpus, we were able to detect hate speech in the HASOC test with a macro F1 score of 0.63. However, the ability of DNNs to learn well is also associated with a higher risk of overfitting. especially when there is little training data (e.g., for HASOC). We looked at several techniques to increase the resources needed for this. We considered several options, including using new models, similarly labeled cases, and unlabeled data. Our results have shown that this can significantly increase the final assessment score.