I guess that the existing mod kneser ney class is not as efficient as it could be (profile this)
In any case it cannot work with our novel sequences out of the box.
So we ether need to integrate the new smoothing techniques into the existing class or more likely reimplement kneser ney smoothing for our novel approach.
The novel kneser ney class should be able to do various smoothing steps (potentially at the same time?)
just the standard kneser ney
the generalized version
the pos version
the hirachical version (first pos then skp)
the mixed version (combination of all varients at the same time)
Results should be logged to different files according to the version of the smoother (e.g. mod-kneser-ney-posglm-5.txt or mod-kneserney-originalglm-5.txt) names.
estimates for the probabilities should always be stored as logs of probabilities.
I guess that the existing mod kneser ney class is not as efficient as it could be (profile this)
In any case it cannot work with our novel sequences out of the box.
So we ether need to integrate the new smoothing techniques into the existing class or more likely reimplement kneser ney smoothing for our novel approach.
The novel kneser ney class should be able to do various smoothing steps (potentially at the same time?)
Results should be logged to different files according to the version of the smoother (e.g. mod-kneser-ney-posglm-5.txt or mod-kneserney-originalglm-5.txt) names.
estimates for the probabilities should always be stored as logs of probabilities.