Multilingual Racism Detection: Effects of CNN Architecture Complexity on Classification Performance

This paper investigates the impact of Convolutional Neural Network (CNN) architecture complexity on multilingual racism detection through text classification. In a previous study, we used self-training to annotate a database using the CNN as a classifier. This study explores various CNN models with increasing architectural depth to discern their effect on classification performance across English, French, and Arabic datasets extracted from social media platforms. We systematically evaluate a range of CNN configurations, varying parameters such as filter sizes and depths of convolutional layers. The experimental results demonstrate that while increasing the complexity of CNN architectures leads to marginal improvements in performance for some languages, the benefits are not uniform across all languages. This study provides valuable insights into optimizing CNN architectures for improved performance in multilingual text classification for racism detection.