[ppod]

This is a complicated problem and is I think best thought of as type of overfitting rather than a complete mistake. The independent or output variable, sentiment, does have an obvious generalisation from movies to politicians, unlike, for example, cinematography quality or trustworthiness. You are also overtraining when you test movie sentiment in the 2010s with reviews trained in the 90s as the concept of sentiment might have shifted if you look at it in that much detail.

(I don't disagree with anything you wrote, just expanding.)