map = word=0,answer=1

bisequenceClassifierPropCh = cn.prop
bisequenceClassifierPropEn = en.prop
supFileEN = en.train.BO

# unannotated bitext 
unsupFileEN = en.unsup.BO
unsupFileCH = cn.unsup.BO
unsupAlignFile = unsup.align

# to speed up training, pre-train the CRF baseline models and load them here
# to generate the weights file, use flag "-serializeWeightsTo $lang.weights.gz"
loadWeightsFromEN = en.weights.gz
# to generate the feature index file, use flag "-serializeFeatureIndexTo $lang.FeatureIndex.gz"
loadFeatureIndexFromEN = en.FeatureIndex.gz

# learned monolingual CH model output to
serializeToCH = cn.bilingual.ser.gz

alignmentPruneThreshold=0.9
doFeatureDiscovery = true
learnCHBasedOnEN = true
learnENBasedOnCH = false

readerAndWriter=edu.stanford.nlp.sequences.ColumnDocumentReaderAndWriter
multiThreadGrad = 10

# set tolerance to greater value if running L-BFGS with early termination
# tolerance = 1e-4

# optimization using L-BFGS
sigma = 20
useQN = true
QNsize = 25
useRobustQN = true

# sometimes L-BFGS runs into numerical stability issues and fails to get pass
# the first few iterations before encountering a NaN exception
# when that happens, consider using AdaGrad+L-BFGS optimization
# i.e., run the first few iterations using AdaGrad, then transition into L-BFGS
# To do so, comment out the above block, and uncomment the following block
sigma = 20
QNsize = 25
useQN = false
useRobustQN = true
useAdaGradFOBOS = true
priorType = gaussian
stochasticBatchSize = -1
initRate = 0.05
adaGradEps = 1e-6
useSGDtoQN = true
SGDPasses=5