In order to improve the accuracy of sentiment text classification

different preprocessing methods of the sentiment of English text is studied

and an improved algorithm of Chi-square statistic（ CHI） feature extraction is put forward. CHI is one of the most efficient feature selection methods

but there are two weaknesses

negative correlation phenomenon and tend to choose low-frequency feature words. In order to overcome these two shortcomings

on the basis of taking into account factors of word frequency

concentration information and dispersion information

considering the length of the text is not balanced and the distribution of feature words

word frequency is normalized

CHI feature extraction algorithm is proposed. Using classical naive Bayes and support vector machine（ SVM） classification algorithms experiments is carried out on balanced corpus

imbalanced corpus and mixed-length corpus

and experimental results show that the new method improves accuracy of sentiment text classification.