HR-LLM: Towards Semantically Enhanced Personnel Retrieval in SOEs with Knowledge Graph and LLM Fine-Tuning

Abstract

Personnel information retrieval in the Smart Organizational Management Platform (SOMP) faces challenges including insufficient semantic understanding, lack of domain-specific knowledge, and data compliance constraints. To address this, we introduce HR-LLM, a lightweight LLM fine-tuning framework for government HR. Our approach integrates an organizational behavior knowledge graph and a user behavior-aware query mechanism to bridge semantic gaps and meet compliance requirements. It encodes rules from documents like the Regulations on the Assessment of Party and Government Leading Cadres into a dual-task loss function and employs a behavior-aware algorithm for personalized searches. Tests on a desensitized SOMP dataset show HR-LLM achieves 92.1% accuracy in parsing complex queries and an 82.3% Recall@5 for professional terms, all while responding in under 500ms on an 8GB GPU. A case study confirms its effectiveness in handling intricate requests, such as finding "department-level cadres with provincial commendations and extensive grassroots experience." This work bridges a gap in semantic understanding of government HR using LLMs and offers a scalable and deployable solution for intelligent talent management systems in state-owned enterprises and government agencies, facilitating technological advancement in smart organizational platforms.

Ähnliche Arbeiten