This paper is to introduce a statistical method to extract Chinese compound words from a very large corpus 1 . This method is based on mutual information and context dependency. Experimental results show that this method is efficient and robust compared with other approaches. We also examined the impact of different parameter settings, corpus size and heterogeneousness on the extraction results. We finally present results on information retrieval to show the usefulness of extracted compounds.