With the development of power business, a large amount of data is produced in the link of customer service. However, traditional sentiment analysis methods for dialogues face many problems and challenges in customer service quality detection. In this study, the word graph is constructed according to the arrangement and location of words, and then the discontinuous long-distance semantic modeling of the whole sentence is carried out. Next, according to the relationship among different parts of the document, the self and interaction dependency relationships between sentence contexts are modeled, respectively. Finally, the convolutional neural network (CNN) is applied to the constructed graph for feature extraction and feature aggregation of the neighbor nodes to obtain the final feature representation of the text. In this way, the detection of emotional states is realized in customer dialogues. Experimental results show that the performance of the proposed model is always higher than that of the baseline model, which demonstrates that the fusion of word co-occurrence relationships, as well as sequential context coding and interactive context coding structures, can effectively improve the accuracy of sentiment category detection. This method provides a fine-grained analysis for intelligently and automatically detecting the emotional states in customer dialogues, which is of great significance to effectively improve the quality of customer service.