We study the spelling suggestion problem for keyword search over an XML document, which provides users with alternative queries that may better express users' search intention. In order to return the query candidates more efficiently, we calculate the correlations between pairwise keywords, which consider the distribution of keywords and the structures of the XML document. We use the keyword correlations as the summaries of the XML document, which are built off-line. We propose an approach to generating the query candidates, and rank them based on the summaries. Experiments with real datasets verifies the effectiveness and efficiency of our approach.