Wrapper for training and testing using SVM_HMM C implementation with default parameters. Consult
http://www.cs.cornell.edu/people/tj/svm_light/svm_hmm.html
for parameter settings.
Parameters:
-c Typical SVM parameter C trading-off slack vs. magnitude of the weight-vector.
NOTE: The default value for this parameter is unlikely to work well for your
particular problem. A good value for C must be selected via cross-validation, ideally
exploring values over several orders of magnitude. NOTE: Unlike in V1.01, the value of C is
divided by the number of training examples. So, to get results equivalent to V1.01, multiply
C by the number of training examples. Default value is set to 1.
-e Parameter "-e <EPSILON>": This specifies the precision to which constraints are
required to be satisfied by the solution. The smaller EPSILON, the longer and the more memory
training takes, but the solution is more precise. However, solutions more accurate than 0.5
typically do not improve prediction accuracy.
-t Order of dependencies of transitions in HMM. Can be any number larger than 1. (default 1)
-m Order of dependencies of emissions in HMM. Can be any number
larger than 0. (default 0) UPDATE: according to svm_struct_api.c: must be either 0 or 1;
fails for >1
-b A non-zero value turns on (approximate) beam search to replace
the exact Viterbi algorithm both for finding the most violated constraint, as well as for
computing predictions. The value is the width of the beam used (e.g. 100). (default 0).