|
| 基于意图识别的知识图谱增强大语言模型问答方法——以防汛抢险为例 |
| Knowledge graph-augmented large language model question answering based on intent recognition: A case study of flood defense and rescue |
| 投稿时间：2025-04-30 修订日期：2025-08-06 |
| DOI： |
| 中文关键词: 防汛抢险 意图识别 知识问答 投票策略 大语言模型 知识图谱 |
| 英文关键词: flood defense and rescue intent recognition knowledge-based question answering voting strategy large language model knowledge graph |
| 基金项目:国家重点研发计划项目课题（2022YFC3005505）；国家自然科学基金项目（52322907，52179141，U23B20149） |
|
| 摘要点击次数: 89 |
| 全文下载次数: 0 |
| 中文摘要: |
| 利用水利专业知识图谱增强大语言模型（LLM）在防汛抢险方面的应用时，用户问句的意图识别仍面临语料匮乏、术语繁多、语义理解困难等挑战，现有方法在小样本意图识别中表现不佳。本文提出一种基于投票策略的多模型融合方法，在小样本条件下准确识别问句意图并提取图谱知识，进而开发水利领域防汛抢险知识问答系统。首先，基于领域实体识别和文本语义表示，构建了基于规则、机器学习和LLM的意图识别单体模型；其次，采用灰狼优化算法，依据单体模型表现分配权重，采用投票策略构建意图识别联合模型。进而，基于联合模型查询防汛抢险知识图谱，基于LLM开发了知识问答系统，实现了自然语言与知识图谱的高效交互。实验结果表明，联合模型在小样本意图识别任务中五折交叉验证的平均F1为0.912，显著超越了以BERT为代表的深度学习模型。所开发防汛抢险知识问答系统实现了准确高效的领域知识检索与重用，为水利知识转化利用和智慧水利建设提供了新路径。 |
| 英文摘要: |
| When enhancing the application of large language model (LLM) in flood defense and rescue with water conservancy knowledge graphs, intent recognition of user queries faces challenges such as limited corpora, an abundance of specialized terminology, and difficulties in semantic understanding. Existing methods perform poorly in low-resource intent recognition scenarios. This study proposes a multi-model ensemble method based on a voting strategy to accurately identify question intent and extract knowledge from knowledge graphs under low-resource conditions, leading to the development of a question-answering system for flood defense and rescue. Firstly, based on domain entity recognition and text semantic representation, three individual intent recognition models were constructed using rule-based methods, machine learning, and LLMs. Secondly, the Grey Wolf Optimization algorithm was used to assign weights to the individual models based on their performance, and a voting strategy was used to construct an intent recognition ensemble model. Finally, the ensemble model was subsequently employed to query the flood defense and rescue knowledge graph, and in combination with an LLM, a question-answering system was developed to facilitate efficient interaction between natural language queries and the knowledge graph. Experimental results show that the ensemble model achieves an average F1 score of 0.912 in five-fold cross-validation in low-resource intent recognition scenarios, markedly outperforming deep learning models such as BERT. The developed system enables accurate and efficient retrieval and reuse of domain knowledge, providing a new pathway for the transformation and utilization of water conservancy knowledge and the advancement of smart water management. |
|
查看/发表评论 下载PDF阅读器 |
| 关闭 |