A Multi-Agent Collaborative Framework for Constructing Knowledge Graphs from Text
Gui Chen, Xianhui Liu · 2024
Recent advancements in large language models (LLMs) have significantly improved natural language understanding and generation, making them valuable tools for knowledge graph construction. However, a single LLM often struggles with the complexity of this task, leading to suboptimal results. To address this challenge, we propose a robust multi-agent collaborative framework for constructing knowledge graphs from text. This framework leverages dynamic interactions among specialized agents, including knowledge graph experts, knowledge extraction experts, data processing experts, and domain-specific experts, to effectively build accurate knowledge graphs from text. Additionally, we introduce a novel prompt construction method tailored for knowledge extraction and a revision mechanism to revise preliminary knowledge graphs. These innovations address common issues in knowledge extraction and enhance the quality of model-generated content. Experimental results on four datasets across two tasks (Named Entity Recognition and Relation Extraction) demonstrate that our approach achieves superior performance in the F1 score compared to baseline methods, highlighting its effectiveness and robustness.