Spammers usually deliver a large number of spam instances generated from a set of templates. To identify spam messages in the same campaigns or to detect new spam instances that are likely to belong to known campaigns, we propose a method to group spam messages based on their HTML struc- tural features. We observe that spam mails tend to have similar structures of the mail bodies, even though the words in the bodies can be signicantly dif- ferent to evade spam detection. Rather than infer the templates and represent them in regular expressions, we extract the HTML tags from the mail bodies as the structural features, and build a ngerprint for each structure. With the ngerprints, we can eciently identify the clusters of similar structures using the simhash algorithm and the Jaccard similarity. The identication is useful to nd new spam instances belonging to known structures with a high recall up to around 95%, while the false-positive rates for normal mails can be less than 5%.