Framing Few-Shot Knowledge Graph Completion with Large Language Models

Abstract

Knowledge Graph Completion (KGC) from text involves identifying known or unknown entities (nodes) as well as relations (edges) among these entities. Recent work has started to explore the use of Large Language Models (LLMs) for entity detection and relation extraction, due to their Natural Language Understanding (NLU) capabilities. However, LLM performance varies across models and depends on the quality of the prompt engineering. We examine specific relation extraction cases and present a set of examples collected from well-known resources in a small corpus. We provide a set of annotations and identify various issues that occur when using different LLMs for this task. As LLMs will remain a focal point of future KGC research, we conclude with suggestions for improving the KGC process.