Abstract: Judging the semantic orientation of words is the basic work of the sentiment analysis of text.This paper proposes a graph—based semi—supervised learning to determine the sentiment polarity of Chinese words.The method constructs the graph,whose nodes consist of Chinese words.The weight of two nodes is based on the semantic similarity of two words in HowNet 2008.With the few labeled sentiment words,the polarity of many unlabeled words can be acquired using label propagation algorithm.Experimental results show that proposed method has higher degree of accuracy compared the traditional methods based on HowNet to judging sentiment polarity of words.