Fine-Grained Balanced Cyberbullying Dataset

Amidst the COVID-19 pandemic, cyberbullying has become an even more serious threat. Our work aims to investigate the viability of an automatic multiclass cyberbullying detection model that is able to classify whether a cyberbully is targeting a victim’s age, ethnicity, gender, religion, or other quality. Previous literature has not yet explored making fine-grained cyberbullying classifications of such magnitude, and existing cyberbullying datasets suffer from quite severe class imbalances. To combat these challenges, we establish a framework for the automatic generation of balanced data by using a semi-supervised online Dynamic Query Expansion (DQE) process to extract more natural data points of a specific class from Twitter. We also propose a Graph Convolutional Network (GCN) classifier, using a graph constructed from the thresholded cosine similarities between tweet embeddings. With our DQE-augmented dataset, which we have made publicly available, we compare our GCN model using eight different tweet embedding methods and six other classification models over two sizes of datasets. Our results show that our proposed GCN model matches or exceeds the performance of the baseline models, as indicated by McNemar statistical tests.