Aimed at the problems that BERT model is poor in domain adaptability, and unable to cope with the problems of uneven number of training data categories and unbalanced classification difficulty, a service text classification method is proposed based on WBBI model. Firstly, the domain adaptability to the BERT model is improved by extending the BERT word list by extracting words from the domain corpus through the TF\|IDF algorithm. Secondly, the service text classification is achieved by the established BERT\|BiLSTM model. Finally, in view of the problems of unbalanced number of categories and unbalanced classification difficulty of the dataset, a zoom loss function is proposedwhich can be dynamically adjusted according to the characteristics of sample unbalance on the basis of the traditional focus loss function. In order to verify the performance of the WBBI model, a large number of comparative experiments are conducted on real datasets obtained from the Internet, and their experimental results show that the WBBI model improves the Macro\|F1 values by 4.29%, 6.59%, 5.3%, and 43% respectively incomparison with the generic text classification models TextCNN, BiLSTM\|attention, RCNN, and Transformer. Compared with the BERT\|based text classification models BERT\|CNN and BERT\|DPCNN, the WBBI model goes further at convergence rate and classifies still better results.