#1 Basic Classifier Configuration and Training
Hi Bill,
What is the optimal Classifier Configuration and
Training Method for the main classifiers for average
purposes?
The (excellent!) crm114-book does not reveal this.
1. Markov: I am doing a SSTTT with pR +/-20 to train
IN. Bill says, it does not work well with DSTTT. But
what about DSTTTR which is very different? However, in
most cases DSTTTR (delta pR 3) would not apply because
after training IN there is mostly a big jump, unlike in
e.g. OSB.
2. OSB: I think SSTTT with pR+/- 10 is a better choice
than pR +/- 20 to train in. Again, what about DSTTTR as
opposed to DSTTT in this classifier? I have many cases
where pR changes only little after training in.
Therefore DSTTTR should be perhaps the training of choice?
3. OSB: Is <unique> superior to "non-<unique"?
4. OSB: <unigram> does not work in the available
Windows binaries.
Regards
Bodo