System design and algorithm optimization of large language model based on knowledge graph in rumour detection

The widespread dissemination of misinformation makes finding new methods that combine unstructured text mining and relational data critically important. This work introduces KG-RumorDetect, a framework that combines Knowledge Graphs (KGs) with Large Language Models (LLMs) for improved rumour detection. Transformer-based language modelling was employed and sophisticated KG integration, graph neural networks, TransE, DistMult and RotatE KG embeddings, multi-modal fusion, and end-to-end joint learning. To further enhance performance and efficiency, the author apply optimisation techniques such as knowledge distillation, model pruning, quantisation, mixed-precision training, dynamic learning rate gradient clipping, and sparse attention with gradual weight adjustment. Extensive experiments confirm that KG-RumorDetect achieves high detection accuracy while significantly reducing computational cost, making it a reliable and efficient option for real-time misinformation detection. Experimental evaluations on benchmark datasets, including Propagation of Health Misinformation in Social Media (PHEME) and Credibility Corpus with Annotations for Events in Social Media (CREDBANK), demonstrate that KG-RumorDetect outperforms standard LLM-based rumour detection models. The proposed model achieves 91.2% accuracy, a 9% improvement over the baseline, with a harmonic mean of the precision and recall (F1-score) of 90.1% and an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.94. The precision-recall analysis highlights that KG-RumorDetect maintains high precision even at high recall levels, reducing false positives by 20% compared to the baseline. Furthermore, training time is reduced by 16% and memory usage by 11%, making the model more suitable for real-time applications.