Sure, I mean naive-wrt-semantics; not naive-wrt-statistics. The naive statistical approach fails, the regularisation approach succeeds. I'd say this does need

Sure, I mean naive-wrt-semantics; not naive-wrt-statistics. The naive statistical approach fails, the regularisation approach succeeds. I'd say this does need explanation, but the paper really omits exactly what needs explaining... it isnt so much the introduction of regularisation... Rather, it's why-it-should-be that our data generating processes encode metadata which can be capture by such regularisation Phrased like this, i htink the answer is somewhat obviuos