The more letters you take to generate the next one, the closer to the original source data you get, but with a big enough corpus of source data, you can still make random names using three or four letters.
The more letters you take to generate the next one, the closer to the original source data you get, but with a big enough corpus of source data, you can still make random names using three or four letters.
Ideally, you would want to go back infinitely many times. My understanding of why Markov models are great is because they reduce the computation required, by not looking back too far, yet they still achieve good results. The more you look back, the more possible combinations (in this case, of letters) there are to consider. As the number of possible combinations of letters increases, the chance of each combination appearing in your training corpus decreases.
N-gram models are often used at the whole word level. That is, instead of the "next letter", they are interested in the "next word". This leads to interesting ways to perform spell checking, based on the context of the surrounding words. For example take the following sentences:
"I did that to"
"I did that too"
The to is a spelling mistake, even though the word is an English word with correct spelling. Imagine how many times the phrase "I did that to" occurs in a large enough corpus, compared to "I did that too".
My understanding is that Google has the largest corpus and largest "N" (they have a 5-gram model). The cool thing is that they have released it under a CC license (http://books.google.com/ngrams/datasets).