From experience, I suspect NB does well on text primarily for two reasons: (1) High dimensionality - this is also the reason why SVMs with linear kernels do so

From experience, I suspect NB does well on text primarily for two reasons: (1) High dimensionality - this is also the reason why SVMs with linear kernels do so well on text (NBs find a linear boundary too - but a different linear boundary than linear-kernel SVMs) (2) this reason applies to certain cases only: text where "token" based estimates are sufficiently discriminatory. For ex take a dataset that has two kinds of documents - one talking about product A and the other talking about product B. And you want to label the documents based on which product they're talking about. Here, just noticing if the term "A" or "B" appears in the document is good enough for classification. You don't have to have powerful models that infer "connotations" of words based on context. NB will do well here, esp if bundled with a feature selection technique that weeds out noisy features (like sthg based on Normalized Mutual Information)