融合阅读理解与扩散模型的实体识别研究

o address the issues of ambiguous entity boundaries and complex semantics in the power sector, a named entity recognition method combining machine reading comprehension and diffusion models is proposed.This approach redefines the NER task as a denoising diffusion process for entity boundaries. During the forward diffusion process, Gaussian noise is incrementally added to the entity boundaries at each time step, while the reverse diffusion process employs a denoising model to progressively refine these boundaries, thereby enhancing the precision of boundary prediction. To further improve model performance, the reverse diffusion process is framed as a question-answering task, where encoding rich prior knowledge helps improve prediction accuracy. Experimental results demonstrate that the MDNER model effectively addresses the intricate issues of entity boundary identification in the power domain, significantly improving recognition accuracy in NER tasks, and providing novel technological support for optimizing equipment management and maintenance decision-making in the electricity sector.