In response to the current issue of low accuracy in sentiment polarity analysis of short texts in power consulting, this paper proposes an improved Latent Dirichlet Allocation (LDA) algorithm-based classification algorithm for power user consulting texts. Based on the analysis of the relationship between power consulting short texts and sentiment, concepts such as sentiment word co-occurrence bags, topic-specific words, and topic relationship words are defined. To improve the quality of semantic analysis, an execution process for the improved LDA algorithm for classifying power user consulting texts is designed. Experiments show that the proposed model demonstrates excellent performance, with an average precision of 90.91% and an average recall rate of 85.03%. The proposed model can fully leverage the advantages of multi-model integration, effectively enhancing the model performance.