| 程环宇,钟秋圆,胡皓,顾彬,相增辉,陈轩,李锦辉,黄祥,胡伟.基于检索增强和测试时扩展的知识图谱问答[J].情报工程,2026,12(4):065-079 |
| 基于检索增强和测试时扩展的知识图谱问答 |
| Retrieval-augmented and Test-time Scaling Knowledge Graph Question Answering |
| |
| DOI： |
| 中文关键词: 知识图谱问答；大语言模型；检索增强；测试时扩展 |
| 英文关键词: Knowledge Graph Question Answering; Large Language Model; Retrieval-augmented Generation; Test-time Scaling |
| 基金项目:国网江苏省电力有限公司科技项目“基于电力信通客服知识图谱的大模型指令微调关键技术研究”（J2025065）。 |
| 作者 | 单位 | | 程环宇 | 1. 国网江苏省电力有限公司 南京 210024 | | 钟秋圆 | 1. 国网江苏省电力有限公司 南京 210024 | | 胡皓 | 1. 国网江苏省电力有限公司 南京 210024 | | 顾彬 | 1. 国网江苏省电力有限公司 南京 210024 | | 相增辉 | 1. 国网江苏省电力有限公司 南京 210024 | | 陈轩 | 1. 国网江苏省电力有限公司 南京 210024 | | 李锦辉 | 1. 国网江苏省电力有限公司 南京 210024 | | 黄祥 | 2. 南京大学 计算机学院 南京 210023 | | 胡伟 | 2. 南京大学 计算机学院 南京 210023 |
|
| 摘要点击次数: 6 |
| 全文下载次数: 6 |
| 中文摘要: |
| [目的/意义]针对基于 Agent 的知识图谱问答中轨迹数据难构造、样例质量要求高、模型容错率低三个问题，提出改进方法。[方法/过程]提出RaTS方法，融合检索增强与测试时扩展。自动将查询转换为轨迹数据，检索最相似样例辅助推理，在每步并行采样多个候选并择优选择。[结果/结论]在 GrailQA 和 GraphQ 数据集上F1分别达到65.8和54.3，相比之前最优方法提升5.3和3.5。方法有效提升了复杂推理的准确性与鲁棒性。 |
| 英文摘要: |
| [Objective/Significance] Focusing on scientific papers as the carrier, this study systematically analyzes the evolutionary characteristics and periodic laws of Shanghai’s global S&T innovation cooperation network, identifies its structural resilience and changes under external shocks, and provides empirical evidence for optimizing the layout of Shanghai’s international S&T cooperation. [Methods/Processes] Taking 160027 internationally co-authored papers published by Shanghai institutions from 2014 to June 2025 in the Web of Science database as research samples, this study employs social network analysis and bibliometric methods to conduct an evolutionary analysis of Shanghai’s global S&T innovation cooperation network from three dimensions: overall network structure, and cooperative disciplines, and cooperative institutions. [Results/Conclusions] Shanghai’s international S&T innovation cooperation network has undergone a four-stage evolution of “initialization-expansionmaturation-adjustment,” with its capability level steadily improving. However, in the fourth stage, the network’s resilience is being tested due to external environmental shocks. The research focus has shifted from local resources to globally cutting-edge fields driven by digital intelligence technologies. It is suggested to take organized collaboration, international S&T organizations,large scientific facilities and digital-intelligent technologies as breakthrough points to comprehensively strengthen the resilience and development level of Shanghai’s international S&T cooperation network. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
