The main reason is that in many cases (but not always) the model obtains enough evidence to make the right decision just from knowing which words appear and don't appear in the document (possibly also using their frequency, but this is not always needed either).
Let's take the textbook example of topic detection from news documents. A 'sports' article is likely to contain at least a few words which are unambiguously related to sports, and the same holds for many topic as long as the topics are sufficiently distinct.
In general tasks which are related to the general semantics of the text work reasonably well with unigrams (single words, unordered) as features, whether with NB or other methods. It's different for tasks which require taking syntax into account, or which require a deeper understanding of the semantics.