|
I'm working with Hidden Markov Models and I have a dataset composed by independent phrases, where each word is an observation. Hence, the best way to adjust my parameters (via Baum-Welch algorithm) is considering each phrase per time and not all phrases concatenated. I would like to know if there is an algorithm that do the training in this way. If not, what are the strategy to avoid transitions created by the concatenation (last word of to first word). |
|
Usually, in HMM what you do is you use your training dataset and maximize using EM.(I think is has a fancier name among HMM people) If you have multiple training observations, like in the case of speech recognition, you have to multiply the likelihood of all of them when doing the HMM, which at the end just incurs in an extra sum among all the observations when you are doing the updates, since you can treat all the training examples as iid. Can you give more details? I don't understand how to combine my inputs in order to adjust the parameters.
(Mar 18 '13 at 15:53)
Gustavo Zeferino
|
Probably you need to explain more how different your situation is from standard BW training setup. There shouldnt' be any differences.
In the standard BW training, all observations are used as if it were a single phrase.