#markov-models

Question 1: Select all that apply

In a Markov model of language, a state can be a longer context, such as the last few words or characters. Which statements about such n-gram models are true? Select all that apply.

Building the transition matrix amounts to counting how often each state follows each other state and normalizing the counts1
In a character-level model whose state is the last 3 characters, each step predicts the next 3 characters at once2
Word n-gram models built on two corpora from completely different domains have the same stationary distribution3
A character-level model over 100 distinct characters whose state is the last 4 characters has a $100 \times 100$ transition matrix4
A longer context typically lowers perplexity on the training text, but too long a context can raise it on held-out text5

Timed test