MAHGA: Multi-Aspect Heterogeneous Graph Analysis for Harmful Speech Detection on Social Networks

Deep neural networks demonstrate high accuracy in detecting harmful social media posts; however, conventional text-based methods often overlook critical contextual relationships among posts, users, and shared information. Graph-based methods, particularly those employing graph neural networks, effectively capture these contextual interactions by modeling relationships within heterogeneous graphs and integrate diverse semantic elements derived through natural language processing. However, existing graph approaches do not adequately differentiate between explicit textual elements (e.g., hashtags, mentions), implicit semantic elements (e.g., entities, topics), and informational or emotional cues from images, thereby obscuring the distinct contribution of each component. This study proposes the multi-aspect heterogeneous graph analysis (MAHGA) framework to address these limitations. MAHGA explicitly models different semantic and emotional aspects of social media posts via specialized heterogeneous graphs. It independently constructs and analyzes distinct graphs for explicit textual elements, implicit semantic elements, and image-based content, enabling a clearer interpretation and more precise understanding of their individual contributions. Furthermore, a correlation-based loss function maintains the independence of learned features, enhancing the model’s ability to recognize nuanced harmful content. Experimental evaluations using COVID-19 vaccine-related social media data indicate that MAHGA significantly outperforms existing methods, offering improvements of approximately 2-3 percentage points in harmful post detection accuracy compared with previous state-of-the-art approaches.