MTBullyGNN: A Graph Neural Network-Based Multitask Framework for Cyberbullying Detection

Cyberbullying is a malady of social media, and its automatic detection is critically important considering its virulence, velocity of spreading, and the scale of the havoc it can wreak. However, the problem is challenging due to its disguised behavior, noise in the content, and, in recent times, introduction of code-mixing. In this work, we propose MTBullyGNN a novel graph neural network (GNN)-based multitask (MT) framework that solves sentiment-aided cyberbullying detection (CD) from code-mixed language. The GNN helps detect unlabelled or noisy label nodes (sentences) accurately by aggregating information from similarly labeled nodes. To connect nodes, we apply cosine similarity between sentences and create a single text graph for a benchmark code-mixed cyberbullying corpus, BullySent. Experimental results illustrate that MTBullyGNN outperforms the state-of-the-art (SOTA) methods for both the single (CD) and MT (CD and sentiment) settings by up to 4.46% and 4.92% in classification accuracy, respectively. Furthermore, another benchmark Hindi-English code-mixed single-task dataset has also been considered to illustrate the robustness of our proposed model. The code will be made publicly available in the camera-ready version.