Power systems are the infrastructure of modern society, and their stability is crucial for the operation of the economy and society. In the face of emergencies such as natural disasters or equipment failures, effective information exchange and rapid decision-making become particularly critical. Existing dialogue systems have many limitations when dealing with power emergency dialogues, especially in intent recognition and slot filling. To improve the performance of dialogue systems, this study proposes a multidimensional span information analysis method based on part-of-speech tagging (MSLA). This method utilizes BERT embeddings and part-of-speech tagging information to enhance the understanding of grammatical structure and semantic relationships in dialogue systems, effectively combining the advantages of pre-trained language models and detailed information from part-of-speech tagging. Experiments show that MSLA achieves better performance than existing models in both intent detection and slot filling tasks, especially on the self-built power emergency dataset, demonstrating its potential for practical application in specific domains.