This is a bit specific. I want to train a support vector machine or naive Bayes classifier to spot advertisements in HTML so I can clean them before LDA analysis. Only a few percentage points of bad data exist, (spam, web menus, headlines; etc. after stripping the text and cleaning with rules). These lines are throwing off my categorization attempts as the good data is lumped together and the bad data is separated by categories, likely because the resulting distance is quite far from the important information. Increasing the number of 'topics' does not appear to help.
Is it possible to use a set such as the Enron spam set to classify advertisements? After finding the root words and removing stopwords, they look fairly similar. Is there a better corpus or approach?