edu.stanford.nlp.parser.shiftreduce

Class ShiftReduceTrainOptions

java.lang.Object

edu.stanford.nlp.parser.lexparser.TrainOptions

edu.stanford.nlp.parser.shiftreduce.ShiftReduceTrainOptions

All Implemented Interfaces:

java.io.Serializable

public class ShiftReduceTrainOptions
extends TrainOptions

See Also:

Serialized Form

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	ShiftReduceTrainOptions.TrainingMethod

Nested classes/interfaces inherited from class edu.stanford.nlp.parser.lexparser.TrainOptions

TrainOptions.TransformMatrixType

Field Summary

Fields

Modifier and Type	Field and Description
float	augmentSubsentences Some training trees will be repeated, with gold transitions given for the first several steps to ensure the parser starts from a good place.
int	averagedModels If set to 0, training outputs the last model produced, regardless of its score.
int	beamSize
boolean	cvAveragedModels Cross-validate over the number of models to average, using the dev set, to figure out which number between 1 and averagedModels we actually want to use
double	decayLearningRate If positive, every 10 iterations, multiply the learning rate by this amount.
static int	DEFAULT_BEAM_SIZE
int	featureFrequencyCutoff How many times a feature must be seen when training.
float	l1Reg If positive, after every iteration, weights are moved by this much back towards 0.
float	l2Reg If positive, after every iteration, weights are scaled by this ratio back towards 0.
boolean	oracleBinaryToShift Does help, but makes the models much bigger for a miniscule gain
boolean	oracleShiftToBinary Does not seem to help...
boolean	retrainAfterCutoff If we cut off features with featureFrequencyCutoff, this retrains with only the existing features
double	retrainShardFeatureDrop Shards 2..n will have this many features dropped.
int	retrainShards If more than one, retrains this many "shards" after the initial training and then averages them together
boolean	saveIntermediateModels Saves intermediate models, but that takes up a lot of space
ShiftReduceTrainOptions.TrainingMethod	trainingMethod

Fields inherited from class edu.stanford.nlp.parser.lexparser.TrainOptions

basicCategoryTagsInDependencyGrammar, batchSize, cheatPCFG, collinsPunc, compactGrammar, debugOutputFrequency, DEFAULT_BATCH_SIZE, DEFAULT_DELTA_MARGIN, DEFAULT_K_BEST, DEFAULT_LEARNING_RATE, DEFAULT_QN_ITERATIONS_PER_BATCH, DEFAULT_REGCOST, DEFAULT_SCALING_FOR_INIT, DEFAULT_STALLED_ITERATION_LIMIT, DEFAULT_TRAINING_ITERATIONS, DEFAULT_UNK_WORD, deleteSplitters, deltaMargin, dvKBest, dvSimplifiedModel, fractionBeforeUnseenCounting, gPA, HSEL_CUT, hSelSplit, learningRate, leftRec, leftToRight, lowercaseWordVectors, markFinalStates, markovFactor, markovOrder, markStrahler, markUnary, markUnaryTags, maxTrainTimeSeconds, noRebinarization, noTagSplit, openClassTypesThreshold, PA, postGPA, postPA, postSplitters, postSplitWithBaseCategory, predictSplits, preTransformer, printAnnotatedPW, printAnnotatedRuleCounts, printAnnotatedStateCounts, printBinarizedPW, printStates, printTreeTransformations, qnEstimates, qnIterationsPerBatch, qnTolerance, randomSeed, regCost, rightRec, ruleDiscount, ruleSmoothing, ruleSmoothingAlpha, scalingForInit, selectivePostSplit, selectivePostSplitCutOff, selectiveSplit, selectiveSplitCutOff, simpleBinarizedLabels, sisterAnnotate, sisterSplitters, smoothing, splitCount, splitPrePreT, splitRecombineRate, splitters, stalledIterationLimit, taggedFiles, tagPA, tagSelectivePostSplit, tagSelectivePostSplitCutOff, tagSelectiveSplit, tagSelectiveSplitCutOff, trainingIterations, trainingThreads, trainLengthLimit, trainTreeFile, trainWordVectors, transformMatrixType, unknownCapsVector, unknownChineseNumberVector, unknownChinesePercentVector, unknownChineseYearVector, unknownDashedWordVectors, unknownNumberVector, unkWord, useContextWords

Constructor Summary

Constructors

Constructor and Description
ShiftReduceTrainOptions()

Method Summary

Methods inherited from class edu.stanford.nlp.parser.lexparser.TrainOptions

compactGrammar, display, outsideFactor, printTrainTree, toString

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, wait, wait, wait

Field Detail

averagedModels

public int averagedModels

If set to 0, training outputs the last model produced, regardless of its score. Otherwise it takes the best k models and averages them together.

cvAveragedModels

public boolean cvAveragedModels

Cross-validate over the number of models to average, using the dev set, to figure out which number between 1 and averagedModels we actually want to use

trainingMethod

public ShiftReduceTrainOptions.TrainingMethod trainingMethod

DEFAULT_BEAM_SIZE

public static final int DEFAULT_BEAM_SIZE

See Also:

Constant Field Values

beamSize

public int beamSize

featureFrequencyCutoff

public int featureFrequencyCutoff

How many times a feature must be seen when training. Less than this and it is filtered.

saveIntermediateModels

public boolean saveIntermediateModels

Saves intermediate models, but that takes up a lot of space

retrainAfterCutoff

public boolean retrainAfterCutoff

If we cut off features with featureFrequencyCutoff, this retrains with only the existing features

oracleShiftToBinary

public boolean oracleShiftToBinary

Does not seem to help... perhaps there is a logic bug in how to compensate for missed binary transitions

oracleBinaryToShift

public boolean oracleBinaryToShift

Does help, but makes the models much bigger for a miniscule gain

decayLearningRate

public double decayLearningRate

If positive, every 10 iterations, multiply the learning rate by this amount.

l1Reg

public float l1Reg

If positive, after every iteration, weights are moved by this much back towards 0.

l2Reg

public float l2Reg

If positive, after every iteration, weights are scaled by this ratio back towards 0.

retrainShards

public int retrainShards

If more than one, retrains this many "shards" after the initial training and then averages them together

retrainShardFeatureDrop

public double retrainShardFeatureDrop

Shards 2..n will have this many features dropped. Shard 1 will be the full model, which guarantees that every feature appears at least once.

augmentSubsentences

public float augmentSubsentences

Some training trees will be repeated, with gold transitions given for the first several steps to ensure the parser starts from a good place. For some datasets, such as the English training set, 0.5 is excessively large.

Constructor Detail

ShiftReduceTrainOptions

public ShiftReduceTrainOptions()