SVllmNER: Self-Verification Large Language Model for Zero-Shot Named Entity Recognition

Recently, Large Language Models (LLMs) are increasingly being used for named entity recognition (NER) due to their excellent natural language processing capabilities. However, when tasked with structured knowledge extraction, the performance of these models still demonstrated inconsistent, especially in zero-shot settings. To address this issue, this paper introduces Self-Verification Large Language Model for Zero-Shot Named Entity Recognition (SVllmNER), which consists of two prompt-based methods by leveraging self-verification ability of LLMs. First, we propose an Entity Correction Mechanism (ECM), which helps to remove the inconsistent entity and modify the incorrectly labeled entity. Second, inspired by the biphasic system in physics, we propose Biphasic Filtering (BF) method to gradually refine predicted information. Regarding labels and entities as two phase state in a system, we use phase threshold at first phase to achieve coarse entity filtering. Then we select the most frequent label for the entity as the final result. Tested on six datasets, empirical results show SVllmNER achieves a significant improvement in LLMs' performance on zero-shot NER. We also explore the combination with other existed strategies. Surprisingly, SVllmNER can help these methods achieve better.