PipeMoE-NER: MoE-LoRA Expert Adaptation and Three-Stage Chain-of-Thought for Pipeline Safety NER

Accurately identifying key entities such as units, risks, consequences, and measures within unstructured texts—including incident reports, inspection records, and management documents—in enterprise pipeline safety scenarios is a critical foundation for risk assessment and emergency decision-making. However, texts in this domain often suffer from dense terminology, drifting semantic boundaries of entities, and long-tailed category distributions. These issues lead to insufficient recall in traditional sequence labeling models and unstable structured outputs or boundary overruns when using end-to-end extraction with large models. To address these challenges, this paper proposes PipeMoE-NER, a specialized adaptation method based on MoE-LoRA and a three-stage Chain-of-Thought (CoT) reasoning approach for Named Entity Recognition (NER) in pipeline safety. On the modeling side, we employ a parameter-efficient expert adaptation strategy that combines the Mixture of Experts (MoE) architecture with the LoRA technique to enhance the modeling capability for long-tail patterns and professional expressions. On the inference side, we design a three-stage Chain-of-Thought prompting process. This process decomposes the extraction task into "candidate generation," "individual discrimination and typing," and "review, verification, and structured output," while introducing explicit constraints to improve type discrimination consistency and JSON output compliance.We constructed and annotated a private Chinese pipeline safety dataset containing five entity types: UNT, RSK, OUT, PRE, and MIT. We evaluated the model using strict entity-level metrics. Experimental results show that PipeMoE-NER achieves an overall F1-score of 81.50 and a Macro-F1 of 80.54 on the test set. This performance surpasses both ChatGLM3-6B Direct (78.49 F1) and DeBERTa-v3+CRF (66.12 F1), validating the effectiveness and robustness of the proposed method for entity recognition tasks in this domain.