I know a lot of spam has the word "Viagra", and very few legitimate emails have the word "Viagra". So if an email contains the word "Viagra", it's more likely to be spam. Naive Bayes classification just applies that to all words in all your emails.