Word Sense Disambiguation (WSD) is the field that seeks to determine the correct sense of a word in a given context. In this paper, we present a WSD method based on random walks over a dependency tree, whose nodes are word-senses from the WordNet. Besides, our method incorporates prior knowledge about the frequency of use of the word-senses. We observed that our results outperform several graph-based WSD methods in All-Word task of SensEval-2 and SensEval-3, including the baseline, where the nouns and verbs part-of-speech show the better improvement in their F-measure scores.