In the field of Natural Language Processing, Named Entity Recognition can be categorized into three main types: flat NER, nested NER, and discontinuous NER. While flat NER has garnered significant attention from researchers, nested NER remains a major challenge. By conducting a thorough analysis of large datasets containing nested entities, this paper identifies distinct patterns within these entities. Building on these patterns, the paper proposes a set of nesting rules and explores their application to the nested NER task. Currently, approaches for handling nested NER include sequence labeling with merged label layers, cascade models, and models based on reading comprehension. Among these, sequence labeling with merged label layers is favored for its simplicity and ease of implementation. Additionally, ChatGPT, a widely used model, is capable of addressing nested NER tasks; thus, the paper applies the proposed nesting rules to both the sequence labeling model and ChatGPT.

To begin, the sequence labeling model with merged label layers is optimized in this study to better incorporate the nesting rules. The experiments utilize a pipeline model to improve the sequence labeling approach, splitting the model into sequence labeling and text classification tasks. During this process, the practice of labeling specific entity categories is abandoned. Instead, entity types are unified into main and subcategories and embedded into the recognized text as identifiers for the text classification task. The model choices for the two tasks include a BERT+BiLSTM+CRF model for sequence labeling and a BERT model for text classification. Experiments on three nested NER datasets: GENIA, CMeEE, and GermEval 2014, demonstrate that the improved models significantly outperform the original methods and exhibit strong competitiveness against existing models. The F1 scores for these datasets were 79.21, 66.71, and 87.81, respectively.

Additionally, the paper attempts to apply these rules to ChatGPT. In ChatGPT, the nesting rules proposed in this study are provided as prompt instructions, which guide ChatGPT in identifying and labeling entities in sentences according to these rules. When testing the GENIA dataset with ChatGPT, the results show a substantial improvement in performance compared to zero-shot, and few-shot modes. Through the application of these nesting rules, the experiments not only improve the accuracy of nested entity recognition but also highlight the need for and potential of further research in this field.