Abstract: It is very important for the text sentiment analysis to build an emotional dictionary. However,most of current researches in this area focus on the words' polarity discrimination. Researchers rarely study the weight assignment of emotional words,and methods on this already existed mostly need to select benchmark words through artificial ways. Using artificial ways brings great difficulty in practical application. To solve this problem,an automatic weight calculation approach of emotional words based on feature selection technique was proposed. Firstly some related assumptions between the emotional weight of words and the emotional tendency of texts were proposed; Then, centered around sentiment classification,the properties of binary classification was combined to improve information gain( IG) and chi-squarec( CHI); Finally,the improved feature selection methods to calculate the weight of emotional words were usesd. Experimental results show that using the emotion dictionary with the calculated weights in text sentiment classification can greatly improve the classification accuracy.