64 lines
1.5 KiB
Plaintext
64 lines
1.5 KiB
Plaintext
# This is better than Jenny's either with or without distsim turned on
|
|
# And using iob2 is better for optimal CoNLL performance.
|
|
# Features titled "chris2009"
|
|
|
|
trainFile = /u/nlp/data/ner/column_data/conll.4class.train
|
|
# testFile = /u/nlp/data/ner/column_data/conll.4class.testa
|
|
serializeTo = english.conll.4class.distsim.crf.ser.gz
|
|
|
|
wordFunction = edu.stanford.nlp.process.AmericanizeFunction
|
|
|
|
useDistSim = true
|
|
distSimLexicon = /u/nlp/data/pos_tags_are_useless/egw4-reut.512.clusters
|
|
# right options for egw4-reut.512 (though effect of having or not is small)
|
|
numberEquivalenceDistSim = true
|
|
unknownWordDistSimClass = 0
|
|
|
|
map = word=0,answer=1
|
|
|
|
saveFeatureIndexToDisk = true
|
|
|
|
useTitle = true
|
|
useClassFeature=true
|
|
useWord=true
|
|
# useWordPairs=true
|
|
useNGrams=true
|
|
noMidNGrams=true
|
|
# maxNGramLeng=6 # Having them all helps, which is the default
|
|
usePrev=true
|
|
useNext=true
|
|
# useTags=true
|
|
# useWordTag=true
|
|
useLongSequences=true
|
|
useSequences=true
|
|
usePrevSequences=true
|
|
maxLeft=1
|
|
useTypeSeqs=true
|
|
useTypeSeqs2=true
|
|
useTypeySequences=true
|
|
useOccurrencePatterns=true
|
|
useLastRealWord=true
|
|
useNextRealWord=true
|
|
#useReverse=false
|
|
normalize=true
|
|
# normalizeTimex=true
|
|
# dan2 better than chris2 on CoNLL data...
|
|
wordShape=dan2useLC
|
|
useDisjunctive=true
|
|
# disjunctionWidth 4 is better than 5 on CoNLL data
|
|
disjunctionWidth=4
|
|
#useDisjunctiveShapeInteraction=true
|
|
|
|
type=crf
|
|
|
|
readerAndWriter=edu.stanford.nlp.sequences.ColumnDocumentReaderAndWriter
|
|
|
|
useObservedSequencesOnly=true
|
|
|
|
sigma = 20
|
|
useQN = true
|
|
QNsize = 25
|
|
|
|
# makes it go faster
|
|
featureDiffThresh=0.05
|