关键词: 关键词: 问答系统, 知识图谱, 图神经网络, 深度推理, 跨源门控

Abstract: Abstract: Abstract: To address strong professionalism, heterogeneous information fusion, and complex reasoning challenges in meteorological question answering (QA), this paper proposes a framework combining Graph Neural Network (GNN)-based deep reasoning and cross-source gating (GDR-CSG). Guided by intent classification, the framework constructs a semantic-driven dynamic collaborative architecture: a GNN is introduced to model deep entity relations in the structured knowledge graph, while a semantic retrieval module simultaneously obtains unstructured text and image symbolic features. A cross-source gated fusion mechanism is then applied to adaptively weight multi-source knowledge based on question intent, driving a large language model (LLM) to generate answers through forward inference. Experimental results demonstrate that GDR-CSG achieves 0.392, 0.598, and 0.869 in BLEU-4, ROUGE-L, and BERTScore, respectively. Notably, on the complex question subset, its BERTScore reaches 0.863, representing an 8.01% relative improvement over the suboptimal graph-augmented retrieval-generation (Graph-RAG) baseline, while its BLEU-4 relatively increases by 13.23%, validating its robustness in complex multi-hop reasoning. Ablation studies indicate that removing either the GNN reasoning or the gated fusion causes performance declines. Specifically, synergistically removing all core modules degrades the system into a pure LLM backbone, resulting in relative decreases of 25.32% in BERTScore and 44.69% in Exact Match (EM), verifying the necessity and synergy of the proposed modules. This research shows that the proposed method effectively enhances reasoning quality in complex meteorological scenarios, providing a solid technical reference for specialized domain QA systems.

Key words: Keywords: Keywords: Questionanswering system, Knowledge Graph (KG), Graph Neural Network (GNN), deep reasoning, Cross-Source Gating