Detection of racism and xenophobia with deep learning models on social media data

One of the most notable innovations of the digital age is the ability of social media platforms to allow individuals to instantly share their thoughts, feelings, and experiences with millions of people. These interactions can also lead to the rapid spread of racist and xenophobic content that may threaten social cohesion. This study aims to develop a deep learning model that can automatically detect such harmful messages with high accuracy on social media platforms like Twitter. Our hypothesis is that a model adopting a hybrid approach will outperform current text classification techniques. During the research process, tweets collected using Python from January 1, 2020, to October 1, 2022, were categorized into those containing discriminatory language against ethnic and national identities and those that did not. Text preprocessing, tokenization, and conversion into numerical vectors were performed, followed by the application of advanced modeling with bidirectional GRU layers and multi- headed attention mechanisms. The evaluations conducted using the cross-validation method determined that the model effectively classifies messages containing hate speech with an average accuracy rate of 62.92%. These findings provide valuable insights into the model's capacity to classify harmful content in social media texts and offer strategic guidance for social media platforms and related institutions in the automatic detection of racism and xenophobia.