Abstract: With the wide application of large language models in general-domain text processing, verify ⁃ ing the reliability of entity recognition results has become a critical issue. Targeting at power station oper ⁃ ation and maintenance logs, this paper proposes an entity recognition result verification and original text backtracking method integrating a hybrid matching strategy. This method first leverages prompt engineer ⁃ ing to guide large language models to extract structured entities from unstructured logs, and then designs a three level matching strategy together with similarity based method for reliability verification. Experi ⁃ ments demonstrate that the proposed method can achieve an entity verification accuracy of 82 % on the dataset of power station operation and maintenance logs and can effectively identify entities.

Key words: large language model, entity recognition, power station operation and maintenance log, reli ? ability verification

摘要： 随着大语言模型在通用领域文本处理中的广泛应用，如何验证其实体识别结果的可靠性成为关键问题。 针对电站运维日志，提出了一种融合混合匹配策略的实体识别结果验证与原文回溯方法。该方法首先利用提示 工程引导大语言模型从非结构化日志中抽取结构化实体，然后设计三级匹配策略与相似性方法进行可靠性验 证。试验表明，本研究方法在自建的电站运维日志数据集上，实体验证准确率达到82%，能够有效识别实体。

关键词: 大语言模型, 实体识别, 电站运维日志, 可靠性验证