A dual-expert social bot detection model based on R-GCN and LLM fusion reasoning

With the rapid development of social media platforms, automated accounts have become an important factor affecting information dissemination, public opinion governance and platform ecology. This paper takes TwiBot-20 as the research object, focusing on the study of social bot detection based on textual semantic information and social topological information. The research objective of this paper is to combine detection performance with readable expert evidence for social bot detection through a data-driven dual-expert model. Firstly, a textual semantic expert based on supervised fine-tuning large language model is constructed. User description, follower count, following count, tweet count, verification status and active years are organized as user profile input, and the textual semantic expert outputs the semantic-side prediction label and evidence summary. Secondly, a topological structure expert based on relational graph convolutional network is built. The construction of this expert includes the construction of user nodes, the construction of Following and Followers relation edges, and the construction of node features by combining account attributes and textual semantic embeddings. Thirdly, a dual-expert fusion reasoning model based on large language model is constructed. The outputs of textual semantic expert and topological structure expert are organized as structured evidence, and user-specific decision weights computed by a model fitted on the development split are used to guide the large language model to generate the final detection result and explanation report. The proposed model achieves 82.76% accuracy and 84.14% bot F1 on TwiBot-20, improving on the stronger single expert by 2.20 and 1.54% points. Its accuracy matches stacking, while the calibrated ensemble achieves a slightly higher bot F1. The model also produces readable reports of expert disagreement and decision cues, illustrated through correct and incorrect arbitration cases.