News Event Source Analysis with Multimodal Transformer Integration

To address the limitations of current news event tracing methods, including reliance on single-modal information, fragmented evidence, and insufficient temporal dynamic modeling, this paper proposes a hybrid framework that integrates a multimodal Transformer with a Temporal Graph Attention Network (T-GAT) to improve the accuracy and robustness of event source tracing. The framework first extracts enhanced representations from text, images, and audio, and employs a two-stage cross-attention mechanism to achieve deep cross-modal fusion and generate unified event embeddings. Subsequently, temporal embeddings and multi-source propagation relationships are incorporated into T-GAT for structured reasoning to identify event origins and reconstruct dissemination paths. From an engineering perspective, the proposed multimodal fusion and temporal reasoning strategy also provides a generalizable methodology for heterogeneous information analysis in intelligent communication environments, offering potential reference value for electromagnetic information processing and propagation-aware decision support. Furthermore, the framework demonstrates transferable applicability to industrial scenarios. Experimental results show that the proposed method achieves a Top-1 tracing accuracy of 0.92 and a Path-IoU of 0.66 while maintaining superior robustness in propagation path reconstruction. Although the model requires higher computational resources than lightweight baselines, its advantages in tracing accuracy and interpretability provide a reliable solution for complex event attribution tasks.