AgREE: Agentic Reasoning for Knowledge Graph Completion in Emerging Companies

Open-domain Knowledge Graph Completion (KGC) faces significant challenges in an ever-changing world, especially considering the continuous emergence of new entities in everyday affairs. Existing KGC methods mainly rely on the knowledge of pre-trained language models, pre-constructed queries, or single-step retrieval, which usually require extensive monitoring and training data. However, they often fail to capture complete and up-to-date information about unpopular and/or emerging businesses. To this end, we present Agenttic Reasoning for Emerging Entities (AgREE), a novel agent-based framework that combines iterative retrieval operations and multi-step reasoning to dynamically construct three knowledge-rich graphs. Experiments show that, despite requiring no training efforts, AgREE outperforms existing methods for constructing three graphs, especially for emerging entities that were not detected during the training of language models, which are more efficient than previous methods up to 13.7%. In addition, we propose a new evaluation method that addresses the fundamental weaknesses of the existing setup and a new KGC benchmark for emerging businesses. Our work demonstrates the effectiveness of combining agent-based reasoning and strategic information retrieval to maintain up-to-date information graphs in dynamic information environments.

• † Sapienza University of Rome