GL-NER: Generation-Aware Large Language Models for Few-Shot Named Entity Recognition

Xingyu Zhu, Feifei Dai, Xiaoyan Gu, Bo Li, Meiou Zhang, Weiping Wang

Abstract: Nowadays, prompt-based methods are widely used in the realm of few-shot named entity recognition. By guiding Pre-trained Language Models to learn token features from training samples with prompts, they obtain several achievements. Nevertheless, the efficacy of prompt-based methods is curtailed by the limited learning capacity inherent in these models. In response to this constraint, we introduce the Generation-aware Large Language Model for few-shot named entity recognition (GL-NER). Our innovation lies in a novel prompt template for Large Language Model (LLM), incorporating label-injected instructions and directly output entity name or “does not exist” to accurately represent recognition results. Furthermore, we enhance few-shot recognition performance through a masking-based loss optimization strategy, facilitating LLM in precise generation. Demonstrating superior performance on widely-used corpora, our approach establishes itself as a state-of-the-art solution in few-shot named entity recognition, attesting to its effectiveness and proficiency.