I was trying to train a MMIE model on a 60 hours training data. I tried all
possible settings "rand", "best", and "ci". In all cases, I couldn't
outperform the accuracy of the conventional Maximum Likelihood Estimate
training. Do you have any explanation? Were you able to see an accuracy
improvement using the MMIE on Sphinx?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I was trying to train a MMIE model on a 60 hours training data. I tried all
possible settings "rand", "best", and "ci". In all cases, I couldn't
outperform the accuracy of the conventional Maximum Likelihood Estimate
training. Do you have any explanation? Were you able to see an accuracy
improvement using the MMIE on Sphinx?
I see this as well. Currently MMIE doesn't really work.