BotSCG: Twitter Bot Detection With Semantic Consistency and Dual-Modulated Graph Learning

Twitter bot detection has emerged as a critical yet challenging task for preventing public opinion manipulation and ensuring the authenticity of online interactions. Current state-of-the-art detection approaches leverage multiple modalities, including user metadata, textual content, and social network relationships. However, these approaches often fail to adequately account for the influence of tweet sequence positions or to exploit the semantic and structural relevance among neighboring nodes, which limits their effectiveness in detecting advanced Twitter bots that deliberately mimic genuine user behavior. To address this issue, a novel multimodal detection framework is proposed, which jointly captures semantic relationships and social network structures for Twitter bot detection. Specifically, BotSCG first introduces a multi-head attention mechanism with sequential relative position bias to extract the semantic consistency of individual users. Then, a dual-modulation mechanism over the connection and relevance matrices to construct a structurally constrained semantic graph, which models the semantic and structural interactions among neighboring nodes and yields user semantic relationship representations. Finally, the semantic consistency features are fused with user relevance representations to measure the overall information consistency of users. Extensive experimental results demonstrate that BotSCG consistently outperforms baselines in terms of accuracy and F1-score on two widely used Twitter bot detection benchmark datasets.