摘要
Entity Resolution(ER)is vital for data integration and knowledge graph construction.Despite the advancements made by deep learning(DL)methods using pre-trained language models(PLMs),these approaches often struggle with unstructured,long-text entities(ULE)in real-world scenarios,where critical information is scattered across the text,and existing DL methods require extensive human labeling and computational resources.To tackle these issues,we propose a Few-shot Uncertainty-calibrated Structural data Enrichment method for ER(FUSER).applies unsupervised pairwise enrichment to extract structural attributes from unstructured entities via Large Language Models(LLMs),and integrates an uncertainty-based calibration module to reduce hallucination issues with minimal additional inference cost.It also implements a lightweight ER pipeline that efficiently performs both blocking and matching tasks with as few as 50 labeled positive samples.was evaluated on six ER benchmark datasets featuring entities,outperforming state-of-the-art methods and significantly boosting the performance of existing ER approaches through its data enrichment component,with a 10 speedup in uncertainty quantification for compared to baseline methods,demonstrating its efficiency and effectiveness in real-world applications.
