make svm faster (cuda)?
 other learning frameworks
target side rule ngram feature template
decoder meta-parameters test
sa-extract -> leave-one-out?
rerank while sgd?