This has been done for the last 20 years (sort of in secret) in the Automatic Speech Recognition community. In brief, if A is the auditory observation you're trying to transcribe and W a possible word sequence, then you use Bayes law to rewrite
P(W|A) = P(A|W)P(W)
The first is known as the auditory model which is responsible for transcribing sounds into phonemes and potential words. The second is the language model responsible just for ranking likely word sequences.
The first tends to be a lot harder than the second, it's models are fuzzier. So to compensate, the real, in practice model is
P(W|A) = P(W|A)P(W)^a
where a is a positive number. I forget what the typical value of the fudge factor is, but generally you use cross validation to optimize it.
Exactly one class told me about the existence of it. Most others pretend like it's not there.
The first tends to be a lot harder than the second, it's models are fuzzier. So to compensate, the real, in practice model is
where a is a positive number. I forget what the typical value of the fudge factor is, but generally you use cross validation to optimize it.Exactly one class told me about the existence of it. Most others pretend like it's not there.