The smileys with obvious sentiment orientation are easily annotated manually. But the annotations of the smileys with unobvious sentiment orientation are difficult to reach a consensus. A method of automatically determining the sentiment orientation of the microblog smileys with the seed words is proposed. The method automatically annotates the corpus smileys with obvious sentiment orientation using a few seed emotions. Then these smileys are used to generate the labeled smiley set (LSS). Moreover, a model is built based on the seed emotional words and LSS to determine the smileys with unobvious sentiment orientation. Experimental results show that the presented method is effective.