edu.stanford.nlp.parser.nndep.Config Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of stanford-parser Show documentation
Show all versions of stanford-parser Show documentation
Stanford Parser processes raw text in English, Chinese, German, Arabic, and French, and extracts constituency parse trees.
package edu.stanford.nlp.parser.nndep;
import edu.stanford.nlp.international.Languages;
import edu.stanford.nlp.ling.HasWord;
import edu.stanford.nlp.tagger.maxent.MaxentTagger;
import edu.stanford.nlp.trees.TreebankLanguagePack;
import edu.stanford.nlp.util.PropertiesUtils;
import edu.stanford.nlp.util.ReflectionLoading;
import java.util.List;
import java.util.Properties;
import java.util.function.Function;
/**
* Defines configuration settings for training and testing the
* neural-network dependency parser.
*
* @see DependencyParser
*
* @author Danqi Chen
* @author Jon Gauthier
*/
public class Config
{
/**
* Out-of-vocabulary token string.
*/
public static final String UNKNOWN = "-UNKNOWN-";
/**
* Root token string.
*/
public static final String ROOT = "-ROOT-";
/**
* Non-existent token string.
*/
public static final String NULL = "-NULL-";
/**
* Represent a non-existent token.
*/
public static final int NONEXIST = -1;
/**
* For printing messages.
*/
public static final String SEPARATOR = "###################";
/**
* The language being parsed.
*/
public Languages.Language language = Languages.Language.English;
/**
* Number of threads to use during training. Also indirectly controls
* how mini-batches are partitioned (more threads => more partitions
* => smaller partitions).
*/
public int trainingThreads = 1;
/**
* Refuse to train on words which have a corpus frequency less than
* this number.
*/
public int wordCutOff = 1;
/**
* Model weights will be initialized to random values within the
* range {@code [-initRange, initRange]}.
*/
public double initRange = 0.01;
/**
* Maximum number of iterations for training
*/
public int maxIter = 20000;
/**
* Size of mini-batch for training. A random subset of training
* examples of this size will be used to train the classifier on each
* iteration.
*/
public int batchSize = 10000;
/**
* An epsilon value added to the denominator of the AdaGrad
* expression for numerical stability
*/
public double adaEps = 1e-6;
/**
* Initial global learning rate for AdaGrad training
*/
public double adaAlpha = 0.01;
/**
* Regularization parameter. All weight updates are scaled by this
* single parameter.
*/
public double regParameter = 1e-8;
/**
* Dropout probability. For each training example we randomly choose
* some amount of units to disable in the neural network classifier.
* This probability controls the proportion of units "dropped out."
*/
public double dropProb = 0.5;
/**
* Size of the neural network hidden layer.
*/
public int hiddenSize = 200;
/**
* Dimensionality of the word embeddings used
*/
public int embeddingSize = 50;
/**
* Total number of tokens provided as input to the classifier. (Each
* token is provided in word embedding form.)
*/
// TODO: we can figure this out automatically based on features used.
// Should remove this option once we make feature templates / dynamic features
public static final int numTokens = 48;
/**
* Number of input tokens for which we should compute hidden-layer
* unit activations.
*
* If zero, the parser will skip the pre-computation step.
*/
public int numPreComputed = 100000;
/**
* During training, run a full UAS evaluation after every
* {@code evalPerIter} iterations.
*/
public int evalPerIter = 100;
/**
* During training, clear AdaGrad gradient histories after every
* {@code clearGradientsPerIter} iterations. (If zero, never clear
* gradients.)
*/
public int clearGradientsPerIter = 0;
/**
* Save an intermediate model file whenever we see an improved UAS
* evaluation. (The frequency of these evaluations is configurable as
* well; see {@link #evalPerIter}.)
*/
public boolean saveIntermediate = true;
/**
* Describes language-specific properties necessary for training and
* testing. By default,
* {@link edu.stanford.nlp.trees.PennTreebankLanguagePack} will be
* used.
*/
public TreebankLanguagePack tlp;
// --- Runtime parsing options
/**
* If non-null, when parsing raw text assume sentences have already
* been split and are separated by the given delimiter.
*
* If null, the parser splits sentences automatically.
*/
public String sentenceDelimiter = null;
/**
* Defines a word-escaper to use when parsing raw sentences.
*
* As a command-line option, you should provide the fully qualified
* class name of a valid escaper (that is, a class which implements
* {@code Function, List>}).
*/
public Function, List> escaper = null;
/**
* Path to a tagger file compatible with
* {@link edu.stanford.nlp.tagger.maxent.MaxentTagger}.
*/
public String tagger = MaxentTagger.DEFAULT_JAR_PATH;
public Config(Properties properties) {
setProperties(properties);
}
private void setProperties(Properties props) {
trainingThreads = PropertiesUtils.getInt(props, "trainingThreads", trainingThreads);
wordCutOff = PropertiesUtils.getInt(props, "wordCutOff", wordCutOff);
initRange = PropertiesUtils.getDouble(props, "initRange", initRange);
maxIter = PropertiesUtils.getInt(props, "maxIter", maxIter);
batchSize = PropertiesUtils.getInt(props, "batchSize", batchSize);
adaEps = PropertiesUtils.getDouble(props, "adaEps", adaEps);
adaAlpha = PropertiesUtils.getDouble(props, "adaAlpha", adaAlpha);
regParameter = PropertiesUtils.getDouble(props, "regParameter", regParameter);
dropProb = PropertiesUtils.getDouble(props, "dropProb", dropProb);
hiddenSize = PropertiesUtils.getInt(props, "hiddenSize", hiddenSize);
embeddingSize = PropertiesUtils.getInt(props, "embeddingSize", embeddingSize);
numPreComputed = PropertiesUtils.getInt(props, "numPreComputed", numPreComputed);
evalPerIter = PropertiesUtils.getInt(props, "evalPerIter", evalPerIter);
clearGradientsPerIter = PropertiesUtils.getInt(props, "clearGradientsPerIter", clearGradientsPerIter);
saveIntermediate = PropertiesUtils.getBool(props, "saveIntermediate", saveIntermediate);
// Runtime parsing options
sentenceDelimiter = PropertiesUtils.getString(props, "sentenceDelimiter", sentenceDelimiter);
tagger = PropertiesUtils.getString(props, "tagger.model", tagger);
String escaperClass = props.getProperty("escaper");
escaper = escaperClass != null ? ReflectionLoading.loadByReflection(escaperClass) : null;
// Language options
language = props.containsKey("language")
? getLanguage(props.getProperty("language"))
: language;
tlp = Languages.getLanguageParams(language).treebankLanguagePack();
}
/**
* Get the {@link edu.stanford.nlp.international.Languages.Language}
* object corresponding to the given language string.
*
* @return A {@link edu.stanford.nlp.international.Languages.Language}
* or {@code null} if no instance matches the given string.
*/
private Languages.Language getLanguage(String languageStr) {
for (Languages.Language l : Languages.Language.values()) {
if (l.name().equalsIgnoreCase(languageStr))
return l;
}
return null;
}
public void printParameters() {
System.err.printf("language = %s%n", language);
System.err.printf("trainingThreads = %d%n", trainingThreads);
System.err.printf("wordCutOff = %d%n", wordCutOff);
System.err.printf("initRange = %.2g%n", initRange);
System.err.printf("maxIter = %d%n", maxIter);
System.err.printf("batchSize = %d%n", batchSize);
System.err.printf("adaEps = %.2g%n", adaEps);
System.err.printf("adaAlpha = %.2g%n", adaAlpha);
System.err.printf("regParameter = %.2g%n", regParameter);
System.err.printf("dropProb = %.2g%n", dropProb);
System.err.printf("hiddenSize = %d%n", hiddenSize);
System.err.printf("embeddingSize = %d%n", embeddingSize);
System.err.printf("numPreComputed = %d%n", numPreComputed);
System.err.printf("evalPerIter = %d%n", evalPerIter);
System.err.printf("clearGradientsPerIter = %d%n", clearGradientsPerIter);
System.err.printf("saveItermediate = %b%n", saveIntermediate);
}
}