The paper "LinkNER: Linking Local Named Entity Recognition Models to Large Language Models using Uncertainty" addresses the challenges of Named Entity Recognition (NER) in web-related applications, where fine-tuned NER models often struggle with unseen entities due to limited training data and lack of external knowledge. The authors propose a framework called LinkNER, which combines a fine-tuned local NER model with a Large Language Model (LLM) to enhance entity recognition performance. The key contribution is the RDC (Recognition-Detection-Classification) strategy, which leverages uncertainty estimation to link the two models. The local model, equipped with uncertainty estimation methods like Confidence, Entropy, Monte Carlo Dropout, and Evidential-based learning, identifies uncertain entities and sends them to the LLM for classification. The LLM, such as GPT-3.5 or Llama 2-Chat, then refines the classification of these entities. Extensive experiments on various datasets, including standard NER benchmarks and noisy social media datasets, demonstrate that LinkNER significantly improves NER performance, outperforming state-of-the-art (SOTA) models in robustness tests. The study also analyzes the impact of uncertainty estimation methods, LLMs, and in-context learning on LinkNER's performance, providing insights into their effectiveness in different scenarios.The paper "LinkNER: Linking Local Named Entity Recognition Models to Large Language Models using Uncertainty" addresses the challenges of Named Entity Recognition (NER) in web-related applications, where fine-tuned NER models often struggle with unseen entities due to limited training data and lack of external knowledge. The authors propose a framework called LinkNER, which combines a fine-tuned local NER model with a Large Language Model (LLM) to enhance entity recognition performance. The key contribution is the RDC (Recognition-Detection-Classification) strategy, which leverages uncertainty estimation to link the two models. The local model, equipped with uncertainty estimation methods like Confidence, Entropy, Monte Carlo Dropout, and Evidential-based learning, identifies uncertain entities and sends them to the LLM for classification. The LLM, such as GPT-3.5 or Llama 2-Chat, then refines the classification of these entities. Extensive experiments on various datasets, including standard NER benchmarks and noisy social media datasets, demonstrate that LinkNER significantly improves NER performance, outperforming state-of-the-art (SOTA) models in robustness tests. The study also analyzes the impact of uncertainty estimation methods, LLMs, and in-context learning on LinkNER's performance, providing insights into their effectiveness in different scenarios.