Quote Originally Posted by warty goblin View Post
Not really, no. For movie A you have a bunch of responses subject to completely unknown biases. For movie B you have a bunch of reponses subject to completely unknown biases. There's no reason to think that movie A mean - movie B mean magically gets those biases to cancel out.

Now I suppose there is a not totally compromised analysis you could do, if you could tie reviews to specific people (i.e. person 1 liked A, hated B, person 2 liked A, liked B, etc). Using this you could construct an estimator of P(like A given like B) or similar quantities. However the sample is still hopelessly biased, and I would not generalize it at all.

(This does not imply that reviews are useless, merely that you cannot base valid statistical inference on them. If you're using them to deduce whether or not you are likely to like movie A, given how both you and the reviewer have felt about past movies, this is utterly reasonable.)
These are good ideas, but I think you could do more.