With the development and maturity of text classification technology, more and more enterprises have applied it to the customer complaint classification system, and obtained the certain effect. Given that the CHI-square Statistic methods tend to choose negative words, so an improved CHI statistical method is applied to the text feature selection, which means reducing the weight of negative words in the feature selection algorithm and minimizing the impact on the model. Finally, an experiment is performed on the complaint text of a communications company business. The result shows that the model and method are effective, and can be more accurate for the classification of business complaints, so as to provide data support for the follow-up problem analysis.