融合动态知识图谱构建的跨语言船舶设计问答方法

A Cross-Lingual Question-Answering Method for Ship Design Integrating Dynamic Knowledge Graph Construction

摘要: 针对船舶情报系统中知识图谱构建成本高、多语言知识碎片化及动态更新困难等问题，文章提出融合动态知识图谱构建的跨语言检索增强生成方法。该方法通过迭代式精细化多标签抽取算法，从多语言非结构化文档中自动构建可演化的船舶设计知识图谱；引入路径置信度加权的混合检索机制，将结构化知识路径与语义检索协同融合；并通过跨语言自然语言推理驱动的协同验证策略，提升生成答案的事实可靠性。实验在英、中、法、德四语种文档上进行评估，多语言平均F1值达86.8%，跨语言实体对齐准确率达94.7%；端到端问答事实正确性为92.3%，较基线方法提升7.8个百分点；针对近期新发布规范的查询，动态更新机制使事实正确性较静态图谱版本提升23.7个百分点，验证了本方法在知识时效性维护方面的有效性。

Abstract: To address the challenges of high construction cost, multilingual knowledge fragmentation, and difficulty in dynamic updating of knowledge graphs in maritime intelligence systems, this paper proposes a cross-lingual retrieval-augmented generation method integrating dynamic knowledge graph construction. The method automatically builds an evolvable ship design knowledge graph from multilingual unstructured documents via an iterative multi-label extraction algorithm, introduces a hybrid retrieval mechanism with path confidence weighting, and enhances the factual reliability of generated answers through a cross-lingual NLI-driven collaborative verification strategy. Evaluated on English, Chinese, French, and German documents, the method achieves a multilingual average F1-score of 86.8% and a cross-lingual entity alignment accuracy of 94.7%. The end-to-end question-answering factual correctness reaches 92.3%, outperforming the baseline by 7.8 percentage points. For queries involving recently updated regulations, the dynamic update mechanism improves factual correctness by 23.7 percentage points over the static graph version, verifying the effectiveness of this method in maintaining the timeliness of knowledge.