Automated Construction of Job–Skill Knowledge Graphs Based on Multi-Agent Consensus and Four-Layer Verification

DOI: https://doi.org/10.62517/jbdc.202601318

Author(s)

Lusheng Sun, Hangyu Zhou, Mengfei Chen, Mengfan Song, Qi Wang*, Hengxu Guo, Jingyuan Yang, Qihang Zhao, Ruian Yan

Affiliation(s)

School of Artificial Intelligence and Big Data, Henan University of Technology, Zhengzhou, China *Corresponding Author

Abstract

To construct reliable job–skill knowledge graphs from heterogeneous recruitment texts, this study combines multi-agent consensus with a four-layer verification cascade. Three large language models from different vendors (iFlytek Spark, DeepSeek, Zhipu GLM-4-Flash) first annotate each JD independently under distinct cognitive roles, then converge through peer review and voting; uncertain skills are subsequently checked against knowledge graph lookup, ESCO alignment, web search, and manual review. Evaluated on 3,942 real-world JDs, the method attains an F1-score of 89.0%, improving by 4.2 points over the best single model and 13.5 points over traditional NER; the cascade reduces the hallucination rate from 15.0% to 1.4%. The ESCO alignment rate of 71.5% leaves room for improvement in covering indigenous Chinese technologies, yet the framework offers a practical pathway for occupational skill standardization and intelligent talent services.

Keywords

Knowledge Graph Construction; Multi-agent Consensus; Large Language Model; Entity Extraction; ESCO International Standard; Job Skills

References

[1] He T, Zhang Q, Zheng G, et al. A Survey of Knowledge Graph Construction in the Era of Large Language Models. Control and Decision, 2025, 40(12): 3509–3527. (in Chinese) [2] Yang Y W, Zhai P Y, Jin Y. Job Entity Recognition and Talent Profiling in the Digital Industry Based on BERT and BiGRU. Journal of Wuhan University (Natural Science Edition), 2025(5): 1–12. (in Chinese) [3] Wang Y S, Shi Y W, et al. Competence Graph of “Gold Majors” in Vocational Education: Logic, Construction Mechanism, and Practice Path. Chinese Vocational and Technical Education, 2024(32): 15–23. (in Chinese) [4] Zhang J, Huang W F, Wu C J, et al. A Survey of LLM-Enhanced Knowledge Graph Construction and Reasoning. Computer Science and Exploration, 2025, 19(11): 2855–2872. (in Chinese) [5] Wang Z Q, Song J X, Yu S S, et al. Construction Method of Smart Mine Knowledge Graph Based on Dependency Parsing. Mining Research and Development, 2023, 43(10): 180–186. (in Chinese) [6] Zhang B A, Cao R Q, et al. Design and Implementation of Vertical Domain Knowledge Graph Construction and Application Platform. Frontiers of Data and Computing, 2023, 5(3): 78–91. (in Chinese) [7] Zhu Y, Wang X, Chen J, et al. LLMs for Knowledge Graph Construction and Reasoning: Recent Capabilities and Future Opportunities. World Wide Web, 2024, 27(5): 58. [8] Yu J L, Su Z W, et al. Employee Training Recommendation Based on Knowledge Graph and Collaborative Filtering. Journal of Yunnan Minzu University (Natural Sciences Edition), 2025(5): 1–9. (in Chinese) [9] Wang Y J, Pan H P. Human Resource Job Matching Prediction Based on Knowledge Graph and Attention Mechanism. Journal of Guiyang University (Natural Sciences Edition), 2025(4): 45–50. (in Chinese) [10] Chen J, Saha S, Bansal M. ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs//Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). Bangkok, 2024: 7066–7085. [11] Zhang Y, Sun R, Chen Y, et al. Chain of Agents: Large Language Models Collaborating on Long-Context Tasks//Advances in Neural Information Processing Systems (NeurIPS). 2024, 37. [12] Lu M, Xie Y, Bi Z, et al. CROSSAGENTIE: Cross-Type and Cross-Task Multi-Agent LLM Collaboration for Zero-Shot Information Extraction//Findings of the Association for Computational Linguistics (ACL). Vienna, 2025: 13953–13977. [13] Haji F, Bethany M, Chiang C, et al. Reflective Agreement: Combining Self-Mixture of Agents with a Sequence Tagger for Robust Event Extraction//Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). Suzhou, 2025: 30476–30492. [14] Ye H, Gui H, Zhang A, et al. CooperKGC: Beyond Isolation—Multi-Agent Synergy for Improving Knowledge Graph Construction//Proceedings of the China Conference on Knowledge Graph and Semantic Computing (CCKS). 2024. arXiv:2312.03022. [15] Kobayashi T, Nakatsuji M. Towards Adaptive Knowledge Structuring by Multi-Agent Consensus//CEUR Workshop Proceedings. 2024, 4093: 1–12. [16] Pan S, Luo L, Wang Y, et al. Unifying Large Language Models and Knowledge Graphs: A Roadmap. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(7): 3580–3599. [17] Pitre P, Ramakrishnan N, Wang X. CONSENSAGENT: Towards Efficient and Effective Consensus in Multi-Agent LLM Interactions Through Sycophancy Mitigation//Findings of the Association for Computational Linguistics (ACL). Vienna, 2025: 22112–22133. [18] de Smedt J, le Vrang M, Papantoniou A. ESCO: Towards a Semantic Web for the European Labour Market//Proceedings of the Workshop on Linked Data on the Web (LDOW 2015), CEUR Workshop Proceedings, Vol. 1409. Florence, Italy, 2015. [19] Marconi G, Baer I, Da Silveira M, et al. Evaluating ESCO Skill Classifiers. Statistical Journal of the IAOS, 2025, 41(4): 1038–1057.