Topic description is as important as topic detection. In this paper, we propose a novel method to describe Web topics with topic words. Under the assumption that representative words exist in important sentences and have high probability of occurrence with other representative words, two graphs are built, one of which represents the relationship for sentences, the other for words. Considering a topic cluster contains a set of different Web pages, sentence clusters are also introduced. Experimental results on a real data set show that our method achieves excellent performance in both high precision and efficiency, especially when real Web data contain mass of noises.