Chain of Thought and Reinforcement Learning Based Low-Resource Named Entity Recognition Model

Baoxing Jiang, Chunyan Han, Jingjing Zhu, Xin Xia, Shenggen Ju

Abstract: Deep learning has significantly advanced named entity recognition (NER), but its performance gains largely depend on large-scale annotated data. In most domains, however, annotated resources remain scarce, and building integrated datasets is costly. The emergence of large language models (LLMs) and prompt learning offers new approaches to mitigate this dependence and address NER tasks in low-resource settings. Yet, existing prompt-based methods often lack sufficient contextual semantic information for entities and fail to fully exploit the relations between entities and their types, making it difficult to capture and optimize entity boundaries and quantity information explicitly. To address these challenges, we design an entity knowledge–enhanced prompt framework that integrates the chain-of-thought paradigm with LoRA fine-tuning. This approach enriches contextual information and guides the model to acquire entity knowledge before prediction. Furthermore, by constructing entity-level preference data and applying the DPO algorithm, we incorporate boundary and quantity information to alleviate category confusion and boundary offset issues. Extensive experiments on four benchmark datasets demonstrate the effectiveness of the proposed method.

External IDs:dblp:conf/IEEEwisa/JiangHZXJ25