The measurement of semantic similarity between words is very important in many applicaitons. In this paper, we propose a method based on Laplacian eigenmaps to measure semantic similarity between words. First, we attach semantic features to each word. Second, a similarity matrix ,which semantic features are encoded into, is calculated in the original high-dimensional space. Finally, with the aid of Laplacian eigenmaps, we recalculate the similarities in the target low-dimensional space. The experiment on the Miller-Charles benchmark shows that the similarity measurement in the low-dimensional space achieves a correlation coefficient of 0.812, in contrast with the correlation coefficient of 0.683 calculated in the high-dimensional space, implying a significant improvement of 18.9%.