CALM: Context Augmentation with Large Language Model for Named Entity Recognition

In prior research on Named Entity Recognition (NER), the focus has been on addressing challenges arising from data scarcity and overfitting, particularly in the context of increasingly complex transformer-based architectures. A framework based on information retrieval (IR), using a search engine to increase input samples and mitigate overfitting tendencies, has been proposed. However, the effectiveness of such system is limited, as they were not designed for this specific application. While this approach serves as a solid foundation, we maintain that LLMs offer capabilities surpassing those of search engines, with greater flexibility in terms of semantic analysis and generation. To overcome these challenges, we propose CALM an innovative context augmentation method, designed for adaptability through prompting. In our study, prompts are meticulously defined as pairs comprising specific tasks and their corresponding response strategies. This careful definition of prompts is pivotal in realizing optimal performance. Our findings illustrate that the resultant context enhances the robustness and performances on NER datasets. We achieve state-of-the-art F1 scores on WNUT17 and CoNLL++. We also delve into the qualitative impact of prompting.