XML is widely applied to represent and exchange data on the Internet. However, XML documents from different sources may convey nearly or exactly the same information but may be different on structures. In previous work, we have proposed LAX(Leaf-clustering based Approximate XML join algorithm), in which the two XML document trees are divided into independent subtrees and the approximate similarity between them are determined by the tree similarity degree based on the mean value of the similarity degrees of matched subtrees. Our previous experimental results show that LAX, comparing with the tree edit distance, is more efficient in performance and more effective for measuring the approximate similarity between XML documents. However, because the tree edit distance is extremely time-consuming, we only used bibliography data of very small sizes to compare the performance of LAX with that of the tree edit distance in our previous experiments. Besides, in LAX, the output is oriented to the pair of documents that have larger tree similarity degree than the threshold. Therefore, when LAX is applied to the fragments divided from large XML documents, the hit subtree selected from the output pair of fragment documents that has large tree similarity degree might not be the proper one that should be integrated. In this paper, we propose SLAX (Subtree-class Leaf-clustering based Approximate XML join algorithm) for integrating the fragments divided from large XML documents by using the maximum match value at subtree classes. And we conduct further experiments to evaluate SLAX, comparing with LAX, by using both real large bibliography and bioinformatics data. The experimental results show that SLAX is more effective than LAX for integrating both large bibliography and bioinformatics data at subtree classes.