KRPO

Knowledge Reconstruction-driven Prompt Optimization (KRPO) is a novel framework designed to enhance the performance of Large Language Models (LLMs) in Open-domain Relational Triplet Extraction (ORTE) tasks. ORTE is crucial for mining structured knowledge without predefined schemas, but existing LLM methods often struggle with semantic ambiguity and perpetuate erroneous extraction patterns due to a lack of reflection. KRPO addresses this by introducing a self-evaluation mechanism that provides intrinsic feedback. This mechanism leverages "knowledge restoration" to project structured triplets into semantic consistency scores, effectively identifying errors. Subsequently, a prompt optimizer, powered by a "textual gradient," internalizes these historical experiences to iteratively refine prompts. This continuous optimization process allows LLMs to adapt and improve their guidance for subsequent extraction tasks, making them more robust and accurate in complex ORTE workflows. It is primarily relevant for researchers and engineers developing advanced knowledge graph construction, information extraction, and LLM-powered data structuring systems.