I'm training a Lidstone Model with different n-gram sequences to see witch one is the best (2-gram, 3-gram, 4-gram, etc) in the same text database.
When I give all these models an unseen text sample from the same database and ask for the perplexity of the n-grams contained in these text, witch perplexity should I expect to be higher?
(Asking to check if my results make sense.)
It's impossible to say. Increasing n trades off variance in exchange for less bias. The only way to know whether increasing n reduces perplexity is by already knowing how exactly how the text was generated.
In practice, unigram models tend to underfit on non-trivial text datasets. 10-gram models trained on small datasets tend to overfit. It's difficult and not really useful to hypothesize about 2,3,4-gram models. Empirically evaluating each on a large independent sample is easy and sufficient.
