Reproduce “Unleashing the Power of Large Language Models in Zero-shot Relation Extraction via Self-Prompting” on OpenResearch

About this paper

Siyi Liu, Yang Li, Jiang Li, Shan Yang, Yunshi Lan

A new self-prompting framework enhances zero-shot relation extraction for large language models by enabling them to generate diverse, context-specific synthetic examples as in-context demonstrations, consistently outperforming existing LLM-based prompting methods and traditional baselines across multiple datasets. The approach achieved improved F1 scores and demonstrated resilience to an increasing number of unseen relations.

Key insights

• Large Language Models can effectively generate high-quality, context-specific in-context learning examples that significantly improve their zero-shot relation extraction capabilities.
• Diversification strategies, including relation synonym generation, entity filtering, and sentence rephrasing, are crucial for creating effective synthetic data and preventing performance degradation from unrefined generation.
• The provision of tailored in-context demonstrations guides LLMs more effectively than generic prompts, leading to enhanced performance and resilience when encountering unseen relations.

Summary by alphaXiv.