Please clarify few details about the third problem.
1) Are we supposed to obtain freq(ab) from the frequency dictionary of trig-rams? (should we sum all the occurrences of trig-rams that contain needed bi-grams?)
2)" If the generated text ends with a N−1 -gram that does not occur in the training data, generate the next character from the full character or ngram distribution. " Does it mean that, for example, if there is no freq(abcd) and therefore the way to compute freq(abcde)/freq(abcd), we should go step lower and try to find freq(abcd)/freq(abc) and so on ? (before we get to the uni-gram distribution?
I have read the previous issue about this topic, but still need more clarification (if possible).
Please clarify few details about the third problem. 1) Are we supposed to obtain freq(ab) from the frequency dictionary of trig-rams? (should we sum all the occurrences of trig-rams that contain needed bi-grams?) 2)" If the generated text ends with a N−1 -gram that does not occur in the training data, generate the next character from the full character or ngram distribution. " Does it mean that, for example, if there is no freq(abcd) and therefore the way to compute freq(abcde)/freq(abcd), we should go step lower and try to find freq(abcd)/freq(abc) and so on ? (before we get to the uni-gram distribution?
I have read the previous issue about this topic, but still need more clarification (if possible).