首页|期刊导航|大数据|本体驱动的电信行业多模态数据集构建与治理方法

本体驱动的电信行业多模态数据集构建与治理方法OA

Ontology-driven construction and governance of multimodal datasets for telecommunications

中文摘要英文摘要

针对电信行业多模态数据集构建中领域知识表达不足、跨模态语义错配和质量治理缺乏闭环的问题,提出本体驱动的构建与治理框架OntoDataset.该框架以领域本体为统一语义基础,采用知识建模、语义推理和数据生成3层架构,并设计了OntAlign算法,融合实体、关系与上下文约束完成跨模态语义对齐.基于某大型电信运营商业务数据的实验表明,综合数据质量由86.7%提高至93.0%,标注效率由250条/小时提高至455条/小时,人工干预率由18%降至12%;在下游任务实验中,客户离网预测的AUC达到0.897,客户意图理解的宏平均F1值和准确率分别达到0.856和0.873.结果表明,该框架可提高垂直行业多模态数据集的质量、一致性和构建效率,为行业大模型的数据集建设与持续治理提供技术支撑.

To address insufficient domain knowledge representation,cross-modal semantic mismatches,and the absence of closed-loop quality governance in multimodal dataset construction for the telecommunications industry,an ontology-driven construction and governance framework,OntoDataset,was proposed.A three-layer architecture for knowledge modeling,semantic reasoning,and data generation was established with a domain ontology as the unified semantic foundation.An ontology-guided algorithm,OntAlign,was also designed to align multimodal data by integrating entity,relation,and contextual constraints.Experiments on real business data from a large telecommunications operator showed that,compared with the best-performing baseline,overall data quality was improved from 86.7%to 93.0%,annotation efficiency was increased from 250 to 455 samples per hour,and the manual intervention rate was reduced from 18%to 12%.In downstream-task evaluations,the AUC for customer churn prediction was 0.897,while the macro-averaged F1 score and accuracy for customer intent recognition were 0.856 and 0.873,respectively.The results indicate that the proposed method improves the quality,consistency,and construction efficiency of multimodal datasets in vertical industries and provides technical support for dataset construction and continuous governance for industry-oriented large models.

张亚威;田亮;马宏伟;王虎;王辰阳;查礼

中国联合网络通信有限公司软件研究院,北京 100176北京东方国信科技股份有限公司,北京 100102北京东方国信科技股份有限公司,北京 100102北京东方国信科技股份有限公司,北京 100102北京东方国信科技股份有限公司,北京 100102北京东方国信科技股份有限公司,北京 100102

信息技术与安全科学

多模态数据集领域本体知识图谱跨模态语义对齐数据质量治理电信行业

multimodal datasetdomain ontologyknowledge graphcross-modal semantic alignmentdata quality governancetelecommunications industry

《大数据》 2026 (5)

43-63,21

10.11959/j.issn.2096-0271.BDR26153

评论