Word sense disambiguation is an automatic selection of correct meaning according to the context of the task, and has become one of the most challenging and the most important problems in the field of computational linguistics, and plays a crucial role in a variety of applications in natural language processing. Therefore, in order to improve the accuracy of word sense disambiguation, an improved unsupervised word sense disambiguation method was proposed. Using “HowNet” as a knowledge base, the method used similarity of a high order new relationship between words measurement method to assign appropriate weights to the edges of a graph. Then, the method was used to select the best fit for the target word. The experiments were carried out in the data set Senseval-3, and the experimental results show that the proposed method could achieve an accuracy of 46.1%, which is better than other unsupervised word sense disambiguation methods under the same test set.