Building an Event Knowledge Graph for Telecom Network Operations: AI R&D Center Case Study

The AI R&D Center of China Telecom Beijing Research Institute focuses on developing AI algorithms for network operation scenarios and shares its experience in building an event knowledge graph for telecom network operations.

Network operation faces growing workload, outdated maintenance models, low data utilization, and limited intelligence, with fault data recorded in work orders that are large, semi‑structured, and often lack standardization.

To address these issues, the center leverages knowledge graphs for strong relational representation, using Neo4j for high‑performance graph storage, enabling faster queries, fault diagnosis, optimization, and intelligent Q&A.

The knowledge graph construction pipeline includes ontology building (based on work order fields such as fault cause, resolution, and equipment), heterogeneous data extraction using the UIE model for entities and attributes, and entity disambiguation, with results stored in Neo4j.

Key steps involve: (1) defining the ontology for work orders, cases, and business rules; (2) extracting entities and attributes from annotated work orders and case documents; (3) training and applying the UIE model for extraction; (4) integrating extracted data into a structured knowledge base.

Applications of the knowledge graph include the Zhixing Cloud Brain Knowledge Base platform for intelligent retrieval and case management, a smart assistant that provides guided recommendations using NLP, and dynamic recommendation of work‑order handling steps based on real‑time graph queries.

The outlook discusses how large language models like ChatGPT could further simplify knowledge extraction, enrich the graph, and automate updates, while emphasizing the continued importance of explainable, graph‑based knowledge for trustworthy network operation support.

A Q&A section addresses the relevance of knowledge graphs in the era of large models and explains the ontology construction process based on work‑order data.