edu.stanford.nlp.tagger.maxent
Class MaxentTagger

java.lang.Object
  extended by edu.stanford.nlp.tagger.maxent.MaxentTagger
All Implemented Interfaces:
SentenceProcessor, ListProcessor<Sentence<? extends HasWord>,Sentence<TaggedWord>>, Function<Sentence<? extends HasWord>,Sentence<TaggedWord>>

public class MaxentTagger
extends Object
implements Function<Sentence<? extends HasWord>,Sentence<TaggedWord>>, SentenceProcessor, ListProcessor<Sentence<? extends HasWord>,Sentence<TaggedWord>>

The main class for users to run, train, and test the part of speech tagger. You can tag things through the Java API or from the command line. The two English taggers included in this distribution are:

Using the Java API

A MaxentTagger can be made with a constructor taking as argument the location of parameter files for a trained tagger:
MaxentTagger tagger = new MaxentTagger("models/left3words-wsj-0-18.tagger");

Alternatively, a constructor with no arguments can be used, which reads the parameters from a default location (which has to be set in the source file, and is set to a path that works on the Stanford NLP machines):
MaxentTagger tagger = new MaxentTagger();

To tag a Sentence and get a TaggedSentence:
Sentence taggedSentence = maxentTagger.tagSentence(Sentence sentence)
Sentence taggedSentence = maxentTagger.apply(Sentence sentence)

To tag a list of sentences and get back a list of tagged sentences:
List taggedList = maxentTagger.process(List sentences)

To tag a String of text and to get back a String with tagged words:
String taggedString = maxentTagger.tagString("Here's a tagged string.")

To tag a string of correctly tokenized, whitespace-separated words and get a string of tagged words back:
String taggedString = maxentTagger.tagTokenizedString("Here 's a tagged string .")

The tagString method uses the default tokenizer (PTBTokenizer). If you wish to control tokenization, you may wish to call tokenizeText(Reader, TokenizerFactory) and then to call process() on the result.

Using the command line

Tagging, testing, and training can all also be done via the command line.

Training from the command line

To train a model from the command line, first generate a property file:
java edu.stanford.nlp.tagger.maxent.MaxentTagger -genprops 
This gets you a default properties file with descriptions of each parameter you can set in your trained model. You can modify the properties file , or use the default options. To train, run:
java -mx1g edu.stanford.nlp.tagger.maxent.MaxentTagger -props myPropertiesFile.props 
with the appropriate properties file specified; any argument you give in the properties file can also be specified on the command line. You must have specified a model using -model, either in the properties file or on the command line, as well as a file containing tagged words using -trainFile.

Tagging and Testing from the command line

Usage: For tagging (plain text):
java edu.stanford.nlp.tagger.maxent.MaxentTagger -model <modelFile> -textFile <textfile> 
For testing (evaluating against tagged text):
java edu.stanford.nlp.tagger.maxent.MaxentTagger -model <modelFile> -testFile <testfile> 
You can use the same properties file as for training if you pass it in with the "-props" argument. The most important arguments for tagging (besides "model" and "file") are "tokenize" and "tokenizerFactory". See below for more details. Note that the tagger assumes input has not yet been tokenized and by default tokenizes it using a default English tokenizer. If your input has already been tokenized, use the flag "-tokenized".

Parameters can be defined using a Properties file (specified on the command-line with -prop propFile), or directly on the command line (by preceding their name with a minust sign ("-") to turn them into a flag. The following properties are recognized:

Property NameTypeDefault ValueRelevant Phase(s)Description
modelStringN/AAllPath and filename where you would like to save the model (training) or where the model should be loaded from (testing, tagging).
trainFileStringN/ATrainPath to the file holding the training data; specifying this option puts the tagger in training mode. Only one of 'trainFile','testFile','texFile', and 'convertToSingleFile' may be specified.
testFileStringN/ATestPath to the file holding the test data; specifying this option puts the tagger in testing mode. Only one of 'trainFile','testFile','texFile', and 'convertToSingleFile' may be specified.
textFileStringN/ATagPath to the file holding the text to tag; specifying this option puts the tagger in tagging mode. Only one of 'trainFile','testFile','textFile', and 'convertToSingleFile' may be specified.
convertToSingleFileStringN/AN/AProvided only for backwards compatibility, this option allows you to convert a tagger trained using a previous version of the tagger to the new single-file format. The value of this flag should be the path for the new model file, 'model' should be the path prefix to the old tagger (up to but not including the ".holder"), and you should supply the properties configuration for the old tagger with -props (before these two arguments).
genpropsbooleanN/AN/AUse this option to output a default properties file, containing information about each of the possible configuration options.
delimiterchar/AllDelimiter character that separates word and part of speech tags. For training and testing, this is the delimiter used in the train/test files. For tagging, this is the character that will be inserted between words and tags in the output.
encodingStringUTF-8AllEncoding of the read files (training, testing) and the output text files.
tokenizebooleantrueTag,TestWhether or not the file has been tokenized (so that white space separates all and only those things that should be tagged as separate words).
tokenizerFactoryStringedu.stanford.nlp.process.PTBTokenizerTag,TestFully qualified classname of the tokenizer to use. edu.stanford.nlp.process.PTBTokenizer does basic English tokenization.
tokenizerOptionsStringTag,TestKnown options for the particular tokenizer used. A comma-separated list. For PTBTokenizer, options of interest include americanize=false and asciiQuotes (for German). Note that any choice of tokenizer options that conflicts with the tokenization used in the tagger training data will likely degrade tagger performance.
archStringgenericTrainArchitecture of the model, as a comma-separated list of options, some with a parenthesized integer argument written k here: this determines what features are sed to build your model. Options are 'left3words', 'left5words', 'bidirectional', 'bidirectional5words', generic', 'sighan2005' (Chinese), 'german', 'words(k),' 'naacl2003unknowns', 'naacl2003conjunctions', wordshapes(k), motleyUnknown, suffix(k), prefix(k), prefixsuffix(k), capitalizationsuffix(k), distsim(s), chinesedictionaryfeatures(s), lctagfeatures, unicodeshapes(k). The left3words architectures are faster, but slightly less accurate, than the bidirectional architectures. 'naacl2003unknowns' was our traditional set of unknown word features, but you can now specify features more flexibility via the various other supported keywords. The 'shapes' options map words to equivalence classes, which slightly increase accuracy.
langStringenglishTrainLanguage from which the part of speech tags are drawn. This option determines which tags are considered closed-class (only fixed set of words can be tagged with a closed-class tag, such as prepositions). Defined languages are 'english' (Penn tagset), 'polish' (very rudimentary), 'chinese', 'arabic', 'german', and 'medline'.
openClassTagsStringN/ATrainSpace separated list of tags that should be considered open-class. All tags encountered that are not in this list are considered closed-class. E.g. format: "NN VB"
closedClassTagsStringN/ATrainSpace separated list of tags that should be considered closed-class. All tags encountered that are not in this list are considered open-class.
learnClosedClassTagsbooleanfalseTrainIf true, induce which tags are closed-class by counting as closed-class tags all those tags which have fewer unique word tokens than closedClassTagThreshold.
closedClassTagThresholdintintTrainNumber of unique word tokens that a tag may have and still be considered closed-class; relevant only if learnClosedClassTags is true.
sgmlbooleanfalseTag, TestVery basic tagging of the contents of all sgml fields; for more complex mark-up, consider using the xmlInput option.
xmlInputStringTag, TestGive a space separated list of tags in an XML file whose content you would like tagged. Any internal tags that appear in the content of fields you would like tagged will be discarded; the rest of the XML will be preserved and the original text of specified fields will be replaced with the tagged text.
xmlOutputString""TagIf a path is given, the tagged data be written out to the given file in xml. If non-empty, each word will be written out within a word tag, with the part of speech as an attribute. If original input was XML, this will just appear in the field where the text originally came from. Otherwise, word tags will be surrounded by sentence tags as well. E.g., <sentence id="0"><word id="0" pos="NN">computer</word></sentence>
tagInsideString""TagTags inside elements that match the regular expression given in the String.
searchStringcgTrainSpecify the search method to be used in the optimization method for training. Options are 'cg' (conjugate gradient) or 'iis' (improved iterative scaling).
sigmaSquareddouble0.5TrainSigma-squared smoothing/regularization parameter to be used for conjugate gradient search. Default usually works reasonably well.
iterationsint100TrainNumber of iterations to be used for improved iterative scaling.
rareWordThreshint5TrainWords that appear fewer than this number of times during training are considered rare words and use extra rare word features.
minFeatureThresholdint5TrainFeatures whose history appears fewer than this number of times are discarded.
curWordMinFeatureThresholdint2TrainWords that occur more than this number of times will generate features with all of the tags they've been seen with.
rareWordMinFeatureThreshint10TrainFeatures of rare words whose histories occur fewer than this number of times are discarded.
veryCommonWordThreshint250TrainWords that occur more than this number of times form an equivalence class by themselves. Ignored unless you are using ambiguity classes.
debugbooleanbooleanAllWhether to write debugging information (words, top words, unknown words). Useful for error analysis.
debugPrefixStringN/AAllFile (path) prefix for where to write out the debugging information (relevant only if debug=true).

Author:
Kristina Toutanova, Miler Lee, Joseph Smarr, Anna Rafferty, Michel Galley, Christopher Manning

Field Summary
static String DEFAULT_DISTRIBUTION_PATH
           
static String DEFAULT_NLP_GROUP_MODEL_PATH
           
 
Constructor Summary
MaxentTagger(String modelFile)
          Constructor for a tagger using a model stored in a particular file.
MaxentTagger(String modelFile, TaggerConfig config)
          Constructor for a tagger using a model stored in a particular file, with options taken from the supplied TaggerConfig.
 
Method Summary
 Sentence<TaggedWord> apply(Sentence<? extends HasWord> in)
          Expects a sentence and returns a tagged sentence.
 TestSentence getTestSentence()
           
static void main(String[] args)
          Command-line tagger interface.
 List<Sentence<TaggedWord>> process(List<? extends Sentence<? extends HasWord>> sentences)
          Tags the Words in each Sentence in the given List with their grammatical part-of-speech.
 Sentence<TaggedWord> processSentence(Sentence sentence)
          Returns a new Sentence that is a copy of the given sentence with all the words tagged with their part-of-speech.
static Sentence<TaggedWord> tagSentence(List<? extends HasWord> sentence)
          Returns a new Sentence that is a copy of the given sentence with all the words tagged with their part-of-speech.
static String tagString(String toTag)
          Tags the input string and returns the tagged version.
static String tagTokenizedString(String toTag)
          Tags the tokenized input string and returns the tagged version.
static List<Sentence<? extends HasWord>> tokenizeText(Reader r)
          Reads data from r, tokenizes it with the default (Penn Treebank) tokenizer, and returns a List of Sentence objects, which can then be fed into tagSentence.
protected static List<Sentence<? extends HasWord>> tokenizeText(Reader r, TokenizerFactory tokenizerFactory)
          Reads data from r, tokenizes it with the given tokenizer, and returns a List of Lists of (extends) HasWord objects, which can then be fed into tagSentence.
 
Methods inherited from class java.lang.Object
clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait
 

Field Detail

DEFAULT_NLP_GROUP_MODEL_PATH

public static final String DEFAULT_NLP_GROUP_MODEL_PATH
See Also:
Constant Field Values

DEFAULT_DISTRIBUTION_PATH

public static final String DEFAULT_DISTRIBUTION_PATH
See Also:
Constant Field Values
Constructor Detail

MaxentTagger

public MaxentTagger(String modelFile)
             throws Exception
Constructor for a tagger using a model stored in a particular file. The modelFile is a filename for the model data. The tagger data is loaded when the constructor is called (this can be slow). This constructor first constructs a TaggerConfig object, which loads the tagger options from the modelFile.

The tagger does not support multithreaded operation. Since some of the data for the tagger is static, two different taggers cannot exist at the same time.

Parameters:
modelFile - filename of the trained model
Throws:
Exception - if IO problem

MaxentTagger

public MaxentTagger(String modelFile,
                    TaggerConfig config)
             throws Exception
Constructor for a tagger using a model stored in a particular file, with options taken from the supplied TaggerConfig. The modelFile is a filename for the model data. The tagger data is loaded when the constructor is called (this can be slow). This version assumes that the tagger options in the modelFile have already been loaded into the TaggerConfig (if that is desired).

The tagger does not support multithreaded operation. Since some of the data for the tagger is static, two different taggers cannot exist at the same time.

Parameters:
modelFile - filename of the trained model
config - The configuration for the tagger
Throws:
Exception - if IO problem
Method Detail

getTestSentence

public TestSentence getTestSentence()

tagTokenizedString

public static String tagTokenizedString(String toTag)
                                 throws Exception
Tags the tokenized input string and returns the tagged version. This method requires the input to already be tokenized. The tagger wants input that is whitespace separated tokens, tokenized according to the conventions of the training data. (For instance, for the Penn Treebank, punctuation marks and possessive "'s" should be separated from words.)

Parameters:
toTag - The untagged input String
Returns:
The same string with tags inserted in the form word/tag
Throws:
Exception - If there are IO errors or class initialization problems

tagString

public static String tagString(String toTag)
Tags the input string and returns the tagged version. This method tokenizes the input into words in perhaps multiple sentences and then tags those sentences. The default (PTB English) tokenizer is used.

Note that this method is static and the model used, etc., will depend on what was set up in an earlier call to the class constructor!

Parameters:
toTag - The untagged input String
Returns:
A String of sentences with tags inserted in the form word/tag

apply

public Sentence<TaggedWord> apply(Sentence<? extends HasWord> in)
Expects a sentence and returns a tagged sentence. The input Sentence items

Specified by:
apply in interface Function<Sentence<? extends HasWord>,Sentence<TaggedWord>>
Parameters:
in - This needs to be a Sentence
Returns:
A Sentence of TaggedWord

process

public List<Sentence<TaggedWord>> process(List<? extends Sentence<? extends HasWord>> sentences)
Tags the Words in each Sentence in the given List with their grammatical part-of-speech. The returned List contains Sentences consisting of TaggedWords.

NOTE: The input document must contain sentences as its elements, not words. To turn a Document of words into a Document of sentences, run it through WordToSentenceProcessor.

Specified by:
process in interface ListProcessor<Sentence<? extends HasWord>,Sentence<TaggedWord>>
Parameters:
sentences - A List of Sentence
Returns:
A List of Sentence of TaggedWord (final generification cannot be listed due to lack of complete generification of super classes)

processSentence

public Sentence<TaggedWord> processSentence(Sentence sentence)
Returns a new Sentence that is a copy of the given sentence with all the words tagged with their part-of-speech. Convenience method when you only want to tag a single Sentence instead of a Document of sentences.

Specified by:
processSentence in interface SentenceProcessor
Parameters:
sentence - A sentence. Classes implementing this interface can assume that the sentence passed in is not null.

tagSentence

public static Sentence<TaggedWord> tagSentence(List<? extends HasWord> sentence)
Returns a new Sentence that is a copy of the given sentence with all the words tagged with their part-of-speech. Convenience method when you only want to tag a single Sentence instead of a Document of sentences.

Parameters:
sentence - sentence to tag
Returns:
tagged sentence

tokenizeText

public static List<Sentence<? extends HasWord>> tokenizeText(Reader r)
Reads data from r, tokenizes it with the default (Penn Treebank) tokenizer, and returns a List of Sentence objects, which can then be fed into tagSentence.

Parameters:
r - Reader where untokenized text is read
Returns:
List of tokenized sentences

tokenizeText

protected static List<Sentence<? extends HasWord>> tokenizeText(Reader r,
                                                                TokenizerFactory tokenizerFactory)
Reads data from r, tokenizes it with the given tokenizer, and returns a List of Lists of (extends) HasWord objects, which can then be fed into tagSentence.

Parameters:
r - Reader where untokenized text is read
tokenizerFactory - Tokenizer. This can be null in which case the default English tokenizer (PTBTokenizerFactory) is used.
Returns:
List of tokenized sentences

main

public static void main(String[] args)
                 throws IOException
Command-line tagger interface. Can be used to train or test taggers, or to tag text, taking input from stdin or a file. See class documentation for usage.

Parameters:
args - Command-line arguments
Throws:
IOException - If any file problems


Stanford NLP Group