Application of Multimodal Knowledge Fusion in Construction of Intelligent Knowledge Graphs

This paper studies the application of multimodal knowledge fusion in the construction of intelligent knowledge graphs, and proposes an innovative method that combines generative adversarial network algorithm and Bayesian network algorithm to improve the efficiency and accuracy of knowledge graph construction. By introducing multimodal data sources, the system can achieve effective knowledge fusion between multiple data modes such as images, texts and voices, thereby enriching and expanding the coverage of knowledge graphs. In order to compare the performance of generative adversarial network algorithm (GANs) and Bayesian network algorithm in the construction of multimodal knowledge graphs, this paper conducts detailed model simulation and data analysis. Through multiple groups of experiments, comparative analysis is conducted from three aspects: accuracy, robustness and stability. In terms of accuracy, experimental results show that GANs exhibit stronger generation ability when processing high-dimensional unstructured data, making the coverage of knowledge graphs wider, and the accuracy of fused knowledge reaches 92.5%. In contrast, although the Bayesian network algorithm performs well in processing structured data, its accuracy is slightly lower than GANs in the fusion of high-dimensional unstructured data, reaching 89.8%. In terms of robustness, the Bayesian network algorithm, due to its inherent probabilistic reasoning mechanism, shows higher robustness when dealing with uncertainty and dependence in multimodal data. In the face of data noise and outliers, the reasoning accuracy of the Bayesian network only decreased by $3.2 \%$, while the accuracy of GANs decreased by $6.5 \%$ under high noise conditions. In terms of stability, the Bayesian network algorithm shows a more stable reasoning process, and its reasoning results in multiple experiments have a smaller variance and show higher consistency. However, due to the characteristics of adversarial learning, GANs may produce large fluctuations between different training rounds, and its reasoning result variance is slightly higher.