This paper proposes a robust method for word sense disambiguation (WSD) of Japanese. Four classifiers were combined in order to improve recall and applicability: one used example sentences in a machine readable dictionary (MRD), one used grammatical information in an MRD, and two classifiers were obtained by supervised learning from a sense-tagged corpus. In other words, we combined several classifiers using heterogeneous language resources, an MRD and a word sense tagged corpus. According to our experimental results, the proposed method outperformed the best single classifier for recall and applicability.