Apache OpenNLP Developer Documentation

OpenNLP provides a command line script, serving as a unique entry point to all included tools. The script is located in the bin directory of OpenNLP binary distribution. Included are versions for Windows: opennlp.bat and Linux or compatible systems: opennlp.

OpenNLP script uses JAVA_CMD and JAVA_HOME variables to determine which command to use to execute Java virtual machine.

OpenNLP script uses OPENNLP_HOME variable to determine the location of the binary distribution of OpenNLP. It is recommended to point this variable to the binary distribution of current OpenNLP version and update PATH variable to include $OPENNLP_HOME/bin or %OPENNLP_HOME%\bin.

Such configuration allows calling OpenNLP conveniently. Examples below suppose this configuration has been done.

Apache OpenNLP provides a common command line script to access all its tools:

                
$ opennlp
                 

This script prints current version of the library and lists all available tools:

                
OpenNLP <VERSION>. Usage: opennlp TOOL
where TOOL is one of:
  Doccat                            learnable document categorizer
  DoccatTrainer                     trainer for the learnable document categorizer
  DoccatConverter                   converts leipzig data format to native OpenNLP format
  DictionaryBuilder                 builds a new dictionary
  SimpleTokenizer                   character class tokenizer
  TokenizerME                       learnable tokenizer
  TokenizerTrainer                  trainer for the learnable tokenizer
  TokenizerMEEvaluator              evaluator for the learnable tokenizer
  TokenizerCrossValidator           K-fold cross validator for the learnable tokenizer
  TokenizerConverter                converts foreign data formats (namefinder,conllx,pos) to native OpenNLP format
  DictionaryDetokenizer
  SentenceDetector                  learnable sentence detector
  SentenceDetectorTrainer           trainer for the learnable sentence detector
  SentenceDetectorEvaluator         evaluator for the learnable sentence detector
  SentenceDetectorCrossValidator    K-fold cross validator for the learnable sentence detector
  SentenceDetectorConverter         converts foreign data formats (namefinder,conllx,pos) to native OpenNLP format
  TokenNameFinder                   learnable name finder
  TokenNameFinderTrainer            trainer for the learnable name finder
  TokenNameFinderEvaluator          Measures the performance of the NameFinder model with the reference data
  TokenNameFinderCrossValidator     K-fold cross validator for the learnable Name Finder
  TokenNameFinderConverter          converts foreign data formats (bionlp2004,conll03,conll02,ad) to native OpenNLP format
  CensusDictionaryCreator           Converts 1990 US Census names into a dictionary
  POSTagger                         learnable part of speech tagger
  POSTaggerTrainer                  trains a model for the part-of-speech tagger
  POSTaggerEvaluator                Measures the performance of the POS tagger model with the reference data
  POSTaggerCrossValidator           K-fold cross validator for the learnable POS tagger
  POSTaggerConverter                converts conllx data format to native OpenNLP format
  ChunkerME                         learnable chunker
  ChunkerTrainerME                  trainer for the learnable chunker
  ChunkerEvaluator                  Measures the performance of the Chunker model with the reference data
  ChunkerCrossValidator             K-fold cross validator for the chunker
  ChunkerConverter                  converts ad data format to native OpenNLP format
  Parser                            performs full syntactic parsing
  ParserTrainer                     trains the learnable parser
  BuildModelUpdater                 trains and updates the build model in a parser model
  CheckModelUpdater                 trains and updates the check model in a parser model
  TaggerModelReplacer               replaces the tagger model in a parser model
All tools print help when invoked with help parameter
Example: opennlp SimpleTokenizer help

                

OpenNLP tools have similar command line structure and options. To discover tool options, run it with no parameters:

                
$ opennlp ToolName
                 

The tool will output two blocks of help.

The first block describes the general structure of this tool command line:

                
Usage: opennlp TokenizerTrainer[.namefinder|.conllx|.pos] [-abbDict path] ...  -model modelFile ...
                

The general structure of this tool command line includes the obligatory tool name (TokenizerTrainer), the optional format parameters ([.namefinder|.conllx|.pos]), the optional parameters ([-abbDict path] ...), and the obligatory parameters (-model modelFile ...).

The format parameters enable direct processing of non-native data without conversion. Each format might have its own parameters, which are displayed if the tool is executed without or with help parameter:

                
$ opennlp TokenizerTrainer.conllx help
                

                
Usage: opennlp TokenizerTrainer.conllx [-abbDict path] [-alphaNumOpt isAlphaNumOpt] ...

Arguments description:
        -abbDict path
                abbreviation dictionary in XML format.
        ...
                

To switch the tool to a specific format, add a dot and the format name after the tool name:

                
$ opennlp TokenizerTrainer.conllx -model en-pos.bin ...
                

The second block of the help message describes the individual arguments:

                
Arguments description:
        -type maxent|perceptron|perceptron_sequence
                The type of the token name finder model. One of maxent|perceptron|perceptron_sequence.
        -dict dictionaryPath
                The XML tag dictionary file
        ...
                

Most tools for processing need to be provided at least a model:

                
$ opennlp ToolName lang-model-name.bin
                 

When tool is executed this way, the model is loaded and the tool is waiting for the input from standard input. This input is processed and printed to standard output.

Alternative, or one should say, most commonly used way is to use console input and output redirection options to provide also an input and an output files:

            
$ opennlp ToolName lang-model-name.bin < input.txt > output.txt
                

Most tools for model training need to be provided first a model name, optionally some training options (such as model type, number of iterations), and then the data.

A model name is just a file name.

Training options often include number of iterations, cutoff, abbreviations dictionary or something else. Sometimes it is possible to provide these options via training options file. In this case these options are ignored and the ones from the file are used.

For the data one has to specify the location of the data (filename) and often language and encoding.

A generic example of a command line to launch a tool trainer might be:

                
$ opennlp ToolNameTrainer -model en-model-name.bin -lang en -data input.train -encoding UTF-8
                 

or with a format:

                
$ opennlp ToolNameTrainer.conll03 -model en-model-name.bin -lang en -data input.train \
                                  -types per -encoding UTF-8
                 

Most tools for model evaluation are similar to those for task execution, and need to be provided fist a model name, optionally some evaluation options (such as whether to print misclassified samples), and then the test data. A generic example of a command line to launch an evaluation tool might be:

                
$ opennlp ToolNameEvaluator -model en-model-name.bin -lang en -data input.test -encoding UTF-8
                 

The easiest way to try out the Sentence Detector is the command line tool. The tool is only intended for demonstration and testing. Download the english sentence detector model and start the Sentence Detector Tool with this command:

        
$ opennlp SentenceDetector en-sent.bin
		

Just copy the sample text from above to the console. The Sentence Detector will read it and echo one sentence per line to the console. Usually the input is read from a file and the output is redirected to another file. This can be achieved with the following command.

				
$ opennlp SentenceDetector en-sent.bin < input.txt > output.txt
		

For the english sentence model from the website the input text should not be tokenized.

The Sentence Detector can be easily integrated into an application via its API. To instantiate the Sentence Detector the sentence model must be loaded first.

				
InputStream modelIn = new FileInputStream("en-sent.bin");

try {
  SentenceModel model = new SentenceModel(modelIn);
}
catch (IOException e) {
  e.printStackTrace();
}
finally {
  if (modelIn != null) {
    try {
      modelIn.close();
    }
    catch (IOException e) {
    }
  }
}
		

After the model is loaded the SentenceDetectorME can be instantiated.

				
SentenceDetectorME sentenceDetector = new SentenceDetectorME(model);
		

The Sentence Detector can output an array of Strings, where each String is one sentence.

				
String sentences[] = sentenceDetector.sentDetect("  First sentence. Second sentence. ");
		

The result array now contains two entries. The first String is "First sentence." and the second String is "Second sentence." The whitespace before, between and after the input String is removed. The API also offers a method which simply returns the span of the sentence in the input string.

				
Span sentences[] = sentenceDetector.sentPosDetect("  First sentence. Second sentence. ");
		

The result array again contains two entries. The first span beings at index 2 and ends at 17. The second span begins at 18 and ends at 34. The utility method Span.getCoveredText can be used to create a substring which only covers the chars in the span.

OpenNLP has a command line tool which is used to train the models available from the model download page on various corpora. The data must be converted to the OpenNLP Sentence Detector training format. Which is one sentence per line. An empty line indicates a document boundary. In case the document boundary is unknown, its recommended to have an empty line every few ten sentences. Exactly like the output in the sample above. Usage of the tool:

				
$ opennlp SentenceDetectorTrainer
Usage: opennlp SentenceDetectorTrainer[.namefinder|.conllx|.pos] [-abbDict path] \
               [-params paramsFile] [-iterations num] [-cutoff num] -model modelFile \
               -lang language -data sampleData [-encoding charsetName]

Arguments description:
        -abbDict path
                abbreviation dictionary in XML format.
        -params paramsFile
                training parameters file.
        -iterations num
                number of training iterations, ignored if -params is used.
        -cutoff num
                minimal number of times a feature must be seen, ignored if -params is used.
        -model modelFile
                output model file.
        -lang language
                language which is being processed.
        -data sampleData
                data to be used, usually a file name.
        -encoding charsetName
                encoding for reading and writing text, if absent the system default is used.
	

To train an English sentence detector use the following command:

				
$ opennlp SentenceDetectorTrainer -model en-sent.bin -lang en -data en-sent.train -encoding UTF-8
                        
        

It should produce the following output:

                
Indexing events using cutoff of 5

	Computing event counts...  done. 4883 events
	Indexing...  done.
Sorting and merging events... done. Reduced 4883 events to 2945.
Done indexing.
Incorporating indexed data for training...  
done.
	Number of Event Tokens: 2945
	    Number of Outcomes: 2
	  Number of Predicates: 467
...done.
Computing model parameters...
Performing 100 iterations.
  1:  .. loglikelihood=-3384.6376826743144	0.38951464263772273
  2:  .. loglikelihood=-2191.9266688597672	0.9397911120212984
  3:  .. loglikelihood=-1645.8640771555981	0.9643661683391358
  4:  .. loglikelihood=-1340.386303774519	0.9739913987302887
  5:  .. loglikelihood=-1148.4141548519624	0.9748105672742167

 ...<skipping a bunch of iterations>...

 95:  .. loglikelihood=-288.25556805874436	0.9834118369854598
 96:  .. loglikelihood=-287.2283680343481	0.9834118369854598
 97:  .. loglikelihood=-286.2174830344526	0.9834118369854598
 98:  .. loglikelihood=-285.222486981048	0.9834118369854598
 99:  .. loglikelihood=-284.24296917223916	0.9834118369854598
100:  .. loglikelihood=-283.2785335773966	0.9834118369854598
Wrote sentence detector model.
Path: en-sent.bin

		

The Sentence Detector also offers an API to train a new sentence detection model. Basically three steps are necessary to train it:

The following sample code illustrates these steps:

				
Charset charset = Charset.forName("UTF-8");				
ObjectStream<String> lineStream =
  new PlainTextByLineStream(new FileInputStream("en-sent.train"), charset);
ObjectStream<SentenceSample> sampleStream = new SentenceSampleStream(lineStream);

SentenceModel model;

try {
  model = SentenceDetectorME.train("en", sampleStream, true, null, TrainingParameters.defaultParams());
}
finally {
  sampleStream.close();
}

OutputStream modelOut = null;
try {
  modelOut = new BufferedOutputStream(new FileOutputStream(modelFile));
  model.serialize(modelOut);
} finally {
  if (modelOut != null) 
     modelOut.close();      
}
		

The command shows how the evaluator tool can be run:

				
$ opennlp SentenceDetectorEvaluator -model en-sent.bin -lang en -data en-sent.eval -encoding UTF-8

Loading model ... done
Evaluating ... done

Precision: 0.9465737514518002
Recall: 0.9095982142857143
F-Measure: 0.9277177006260672
                

The en-sent.eval file has the same format as the training data.

The easiest way to try out the tokenizers are the command line tools. The tools are only intended for demonstration and testing.

There are two tools, one for the Simple Tokenizer and one for the learnable tokenizer. A command line tool the for the Whitespace Tokenizer does not exist, because the whitespace separated output would be identical to the input.

The following command shows how to use the Simple Tokenizer Tool.

			
$ opennlp SimpleTokenizer
		    

To use the learnable tokenizer download the english token model from our website.

			
$ opennlp TokenizerME en-token.bin
		    

To test the tokenizer copy the sample from above to the console. The whitespace separated tokens will be written back to the console.

Usually the input is read from a file and written to a file.

			
$ opennlp TokenizerME en-token.bin < article.txt > article-tokenized.txt
		    

It can be done in the same way for the Simple Tokenizer.

Since most text comes truly raw and doesn't have sentence boundaries and such, its possible to create a pipe which first performs sentence boundary detection and tokenization. The following sample illustrates that.

			
$ opennlp SentenceDetector sentdetect.model < article.txt | opennlp TokenizerME tokenize.model | more
Loading model ... Loading model ... done
done
Showa Shell gained 20 to 1,570 and Mitsubishi Oil rose 50 to 1,500.
Sumitomo Metal Mining fell five yen to 692 and Nippon Mining added 15 to 960 .
Among other winners Wednesday was Nippon Shokubai , which was up 80 at 2,410 .
Marubeni advanced 11 to 890 .
London share prices were bolstered largely by continued gains on Wall Street and technical 
    factors affecting demand for London 's blue-chip stocks .
...etc...
		 

Of course this is all on the command line. Many people use the models directly in their Java code by creating SentenceDetector and Tokenizer objects and calling their methods as appropriate. The following section will explain how the Tokenizers can be used directly from java.

The Tokenizers can be integrated into an application by the defined API. The shared instance of the WhitespaceTokenizer can be retrieved from a static field WhitespaceTokenizer.INSTANCE. The shared instance of the SimpleTokenizer can be retrieved in the same way from SimpleTokenizer.INSTANCE. To instantiate the TokenizerME (the learnable tokenizer) a Token Model must be created first. The following code sample shows how a model can be loaded.

			
InputStream modelIn = new FileInputStream("en-token.bin");

try {
  TokenizerModel model = new TokenizerModel(modelIn);
}
catch (IOException e) {
  e.printStackTrace();
}
finally {
  if (modelIn != null) {
    try {
      modelIn.close();
    }
    catch (IOException e) {
    }
  }
}
		 

After the model is loaded the TokenizerME can be instantiated.

			
Tokenizer tokenizer = new TokenizerME(model);
		 

The tokenizer offers two tokenize methods, both expect an input String object which contains the untokenized text. If possible it should be a sentence, but depending on the training of the learnable tokenizer this is not required. The first returns an array of Strings, where each String is one token.

			
String tokens[] = tokenizer.tokenize("An input sample sentence.");
		 

The output will be an array with these tokens.

			
"An", "input", "sample", "sentence", "."
		 

The second method, tokenizePos returns an array of Spans, each Span contain the begin and end character offsets of the token in the input String.

			
Span tokenSpans[] = tokenizer.tokenizePos("An input sample sentence.");		
			

The tokenSpans array now contain 5 elements. To get the text for one span call Span.getCoveredText which takes a span and the input text. The TokenizerME is able to output the probabilities for the detected tokens. The getTokenProbabilities method must be called directly after one of the tokenize methods was called.

			
TokenizerME tokenizer = ...

String tokens[] = tokenizer.tokenize(...);
double tokenProbs[] = tokenizer.getTokenProbabilities();
			

The tokenProbs array now contains one double value per token, the value is between 0 and 1, where 1 is the highest possible probability and 0 the lowest possible probability.

OpenNLP has a command line tool which is used to train the models available from the model download page on various corpora. The data can be converted to the OpenNLP Tokenizer training format or used directly. The OpenNLP format contains one sentence per line. Tokens are either separated by a whitespace or by a special <SPLIT> tag. The following sample shows the sample from above in the correct format.

			    
Pierre Vinken<SPLIT>, 61 years old<SPLIT>, will join the board as a nonexecutive director Nov. 29<SPLIT>.
Mr. Vinken is chairman of Elsevier N.V.<SPLIT>, the Dutch publishing group<SPLIT>.
Rudolph Agnew<SPLIT>, 55 years old and former chairman of Consolidated Gold Fields PLC<SPLIT>,
    was named a nonexecutive director of this British industrial conglomerate<SPLIT>.
			    

Usage of the tool:

			    
$ opennlp TokenizerTrainer
Usage: opennlp TokenizerTrainer[.namefinder|.conllx|.pos] [-abbDict path] \
                [-alphaNumOpt isAlphaNumOpt] [-params paramsFile] [-iterations num] \
                [-cutoff num] -model modelFile -lang language -data sampleData \
                [-encoding charsetName]

Arguments description:
        -abbDict path
                abbreviation dictionary in XML format.
        -alphaNumOpt isAlphaNumOpt
                Optimization flag to skip alpha numeric tokens for further tokenization
        -params paramsFile
                training parameters file.
        -iterations num
                number of training iterations, ignored if -params is used.
        -cutoff num
                minimal number of times a feature must be seen, ignored if -params is used.
        -model modelFile
                output model file.
        -lang language
                language which is being processed.
        -data sampleData
                data to be used, usually a file name.
        -encoding charsetName
                encoding for reading and writing text, if absent the system default is used.
                

To train the english tokenizer use the following command:

			    
$ opennlp TokenizerTrainer -model en-token.bin -alphaNumOpt -lang en -data en-token.train -encoding UTF-8

Indexing events using cutoff of 5

	Computing event counts...  done. 262271 events
	Indexing...  done.
Sorting and merging events... done. Reduced 262271 events to 59060.
Done indexing.
Incorporating indexed data for training...  
done.
	Number of Event Tokens: 59060
	    Number of Outcomes: 2
	  Number of Predicates: 15695
...done.
Computing model parameters...
Performing 100 iterations.
  1:  .. loglikelihood=-181792.40419263614	0.9614292087192255
  2:  .. loglikelihood=-34208.094253153664	0.9629238459456059
  3:  .. loglikelihood=-18784.123872910015	0.9729211388220581
  4:  .. loglikelihood=-13246.88162585859	0.9856103038460219
  5:  .. loglikelihood=-10209.262670265718	0.9894422181636552

 ...<skipping a bunch of iterations>...

 95:  .. loglikelihood=-769.2107474529454	0.999511955191386
 96:  .. loglikelihood=-763.8891914534009	0.999511955191386
 97:  .. loglikelihood=-758.6685383254891	0.9995157680414533
 98:  .. loglikelihood=-753.5458314695236	0.9995157680414533
 99:  .. loglikelihood=-748.5182305519613	0.9995157680414533
100:  .. loglikelihood=-743.5830058068038	0.9995157680414533
Wrote tokenizer model.
Path: en-token.bin
				

The Tokenizer offers an API to train a new tokenization model. Basically three steps are necessary to train it:

The following sample code illustrates these steps:

                    
Charset charset = Charset.forName("UTF-8");
ObjectStream<String> lineStream = new PlainTextByLineStream(new FileInputStream("en-sent.train"),
    charset);
ObjectStream<TokenSample> sampleStream = new TokenSampleStream(lineStream);

TokenizerModel model;

try {
  model = TokenizerME.train("en", sampleStream, true, TrainingParameters.defaultParams());
}
finally {
  sampleStream.close();
}

OutputStream modelOut = null;
try {
  modelOut = new BufferedOutputStream(new FileOutputStream(modelFile));
  model.serialize(modelOut);
} finally {
  if (modelOut != null)
     modelOut.close();
}
                

TODO: Write documentation about the detokenizer api. Any contributions are very welcome. If you want to contribute please contact us on the mailing list or comment on the jira issue OPENNLP-216.

TODO: Write documentation about the detokenizer dictionary. Any contributions are very welcome. If you want to contribute please contact us on the mailing list or comment on the jira issue OPENNLP-217.

The easiest way to try out the Name Finder is the command line tool. The tool is only intended for demonstration and testing. Download the English person model and start the Name Finder Tool with this command:

				
$ opennlp TokenNameFinder en-ner-person.bin
			 

The name finder now reads a tokenized sentence per line from stdin, an empty line indicates a document boundary and resets the adaptive feature generators. Just copy this text to the terminal:

				
Pierre Vinken , 61 years old , will join the board as a nonexecutive director Nov. 29 .
Mr . Vinken is chairman of Elsevier N.V. , the Dutch publishing group .
Rudolph Agnew , 55 years old and former chairman of Consolidated Gold Fields PLC , was named
    a director of this British industrial conglomerate .
			 

the name finder will now output the text with markup for person names:

				
<START:person> Pierre Vinken <END> , 61 years old , will join the board as a nonexecutive director Nov. 29 .
Mr . <START:person> Vinken <END> is chairman of Elsevier N.V. , the Dutch publishing group .
<START:person> Rudolph Agnew <END> , 55 years old and former chairman of Consolidated Gold Fields PLC ,
    was named a director of this British industrial conglomerate .
			 

To use the Name Finder in a production system its strongly recommended to embed it directly into the application instead of using the command line interface. First the name finder model must be loaded into memory from disk or an other source. In the sample below its loaded from disk.

				
InputStream modelIn = new FileInputStream("en-ner-person.bin");

try {
  TokenNameFinderModel model = new TokenNameFinderModel(modelIn);
}
catch (IOException e) {
  e.printStackTrace();
}
finally {
  if (modelIn != null) {
    try {
      modelIn.close();
    }
    catch (IOException e) {
    }
  }
}
			 

There is a number of reasons why the model loading can fail:

After the model is loaded the NameFinderME can be instantiated.

				
NameFinderME nameFinder = new NameFinderME(model);
			

The initialization is now finished and the Name Finder can be used. The NameFinderME class is not thread safe, it must only be called from one thread. To use multiple threads multiple NameFinderME instances sharing the same model instance can be created. The input text should be segmented into documents, sentences and tokens. To perform entity detection an application calls the find method for every sentence in the document. After every document clearAdaptiveData must be called to clear the adaptive data in the feature generators. Not calling clearAdaptiveData can lead to a sharp drop in the detection rate after a few documents. The following code illustrates that:

				
for (String document[][] : documents) {

  for (String[] sentence : document) {
    Span nameSpans[] = nameFinder.find(sentence);
    // do something with the names
  }

  nameFinder.clearAdaptiveData()
}
			 

the following snippet shows a call to find

				
String sentence[] = new String[]{
    "Pierre",
    "Vinken",
    "is",
    "61",
    "years"
    "old",
    "."
    };

Span nameSpans[] = nameFinder.find(sentence);
			

The nameSpans arrays contains now exactly one Span which marks the name Pierre Vinken. The elements between the begin and end offsets are the name tokens. In this case the begin offset is 0 and the end offset is 2. The Span object also knows the type of the entity. In this case its person (defined by the model). It can be retrieved with a call to Span.getType(). Additionally to the statistical Name Finder, OpenNLP also offers a dictionary and a regular expression name finder implementation.

TODO: Explain how to retrieve probs from the name finder for names and for non recognized names

OpenNLP has a command line tool which is used to train the models available from the model download page on various corpora.

The data can be converted to the OpenNLP name finder training format. Which is one sentence per line. Some other formats are available as well. The sentence must be tokenized and contain spans which mark the entities. Documents are separated by empty lines which trigger the reset of the adaptive feature generators. A training file can contain multiple types. If the training file contains multiple types the created model will also be able to detect these multiple types. For now its recommended to only train single type models, since multi type support is still experimental.

Sample sentence of the data:

				
<START:person> Pierre Vinken <END> , 61 years old , will join the board as a nonexecutive director Nov. 29 .
Mr . <START:person> Vinken <END> is chairman of Elsevier N.V. , the Dutch publishing group .
			 

The training data should contain at least 15000 sentences to create a model which performs well. Usage of the tool:

				
$ opennlp TokenNameFinderTrainer
Usage: opennlp TokenNameFinderTrainer[.bionlp2004|.conll03|.conll02|.ad] [-resources resourcesDir] \
               [-type modelType] [-featuregen featuregenFile] [-params paramsFile] \
               [-iterations num] [-cutoff num] -model modelFile -lang language \
               -data sampleData [-encoding charsetName]

Arguments description:
        -resources resourcesDir
                The resources directory
        -type modelType
                The type of the token name finder model
        -featuregen featuregenFile
                The feature generator descriptor file
        -params paramsFile
                training parameters file.
        -iterations num
                number of training iterations, ignored if -params is used.
        -cutoff num
                minimal number of times a feature must be seen, ignored if -params is used.
        -model modelFile
                output model file.
        -lang language
                language which is being processed.
        -data sampleData
                data to be used, usually a file name.
        -encoding charsetName
                encoding for reading and writing text, if absent the system default is used.
			 

Its now assumed that the english person name finder model should be trained from a file called en-ner-person.train which is encoded as UTF-8. The following command will train the name finder and write the model to en-ner-person.bin:

				
$ opennlp TokenNameFinderTrainer -model en-ner-person.bin -lang en -data en-ner-person.train -encoding UTF-8
			 

Additionally its possible to specify the number of iterations, the cutoff and to overwrite all types in the training data with a single type.

To train the name finder from within an application its recommended to use the training API instead of the command line tool. Basically three steps are necessary to train it:

The three steps are illustrated by the following sample code:

				
Charset charset = Charset.forName("UTF-8");
ObjectStream<String> lineStream =
		new PlainTextByLineStream(new FileInputStream("en-ner-person.train"), charset);
ObjectStream<NameSample> sampleStream = new NameSampleDataStream(lineStream);

TokenNameFinderModel model;

try {
  model = NameFinderME.train("en", "person", sampleStream, TrainingParameters.defaultParams(),
            null, Collections.<String, Object>emptyMap());
}
finally {
  sampleStream.close();
}

try {
  modelOut = new BufferedOutputStream(new FileOutputStream(modelFile));
  model.serialize(modelOut);
} finally {
  if (modelOut != null) 
     modelOut.close();      
}
			 

OpenNLP defines a default feature generation which is used when no custom feature generation is specified. Users which want to experiment with the feature generation can provide a custom feature generator. Either via API or via an xml descriptor file.

					
AdaptiveFeatureGenerator featureGenerator = new CachedFeatureGenerator(
         new AdaptiveFeatureGenerator[]{
           new WindowFeatureGenerator(new TokenFeatureGenerator(), 2, 2),
           new WindowFeatureGenerator(new TokenClassFeatureGenerator(true), 2, 2),
           new OutcomePriorFeatureGenerator(),
           new PreviousMapFeatureGenerator(),
           new BigramNameFeatureGenerator(),
           new SentenceFeatureGenerator(true, false)
           });
				

					
public static TokenNameFinderModel train(String languageCode, String type, ObjectStream<NameSample> samples, 
       TrainingParameters trainParams, AdaptiveFeatureGenerator generator, final Map<String, Object> resources) throws IOException
				

					
new NameFinderME(model, featureGenerator, NameFinderME.DEFAULT_BEAM_SIZE);
				 

					
<generators>
  <cache> 
    <generators>
      <window prevLength = "2" nextLength = "2">          
        <tokenclass/>
      </window>
      <window prevLength = "2" nextLength = "2">                
        <token/>
      </window>
      <definition/>
      <prevmap/>
      <bigram/>
      <sentence begin="true" end="false"/>
    </generators>
  </cache> 
</generators>
				 

The following command shows how the tool can be run:

            
$ opennlp TokenNameFinderEvaluator -model en-ner-person.bin -lang en -data en-ner-person.test -encoding UTF-8

Precision: 0.8005071889818507
Recall: 0.7450581122145297
F-Measure: 0.7717879983140168
         

Note: The command line interface does not support cross evaluation in the current version.

The evaluation can be performed on a pre-trained model and a test dataset or via cross validation. In the first case the model must be loaded and a NameSample ObjectStream must be created (see code samples above), assuming these two objects exist the following code shows how to perform the evaluation:

				
TokenNameFinderEvaluator evaluator = new TokenNameFinderEvaluator(new NameFinderME(model));
evaluator.evaluate(sampleStream);

FMeasure result = evaluator.getFMeasure();

System.out.println(result.toString());
			

In the cross validation case all the training arguments must be provided (see the Training API section above). To perform cross validation the ObjectStream must be resettable.

				
FileInputStream sampleDataIn = new FileInputStream("en-ner-person.train");
ObjectStream<NameSample> sampleStream = new PlainTextByLineStream(sampleDataIn.getChannel(), "UTF-8"); 
TokenNameFinderCrossValidator evaluator = new TokenNameFinderCrossValidator("en", 100, 5);
evaluator.evaluate(sampleStream, 10);

FMeasure result = evaluator.getFMeasure();

System.out.println(result.toString());
			

The easiest way to try out the document categorizer is the command line tool. The tool is only intended for demonstration and testing. The following command shows how to use the document categorizer tool.

			
$ opennlp Doccat model
		 

The input is read from standard input and output is written to standard output, unless they are redirected or piped. As with most components in OpenNLP, document categorizer expects input which is segmented into sentences.

To perform classification you will need a maxent model - these are encapsulated in the DoccatModel class of OpenNLP tools.

First you need to grab the bytes from the serialized model on an InputStream - we'll leave it you to do that, since you were the one who serialized it to begin with. Now for the easy part:

				
InputStream is = ...
DoccatModel m = new DoccatModel(is);
				

With the DoccatModel in hand we are just about there:

				
String inputText = ...
DocumentCategorizerME myCategorizer = new DocumentCategorierME(m);
double[] outcomes = myCategorizer.categorize(inputText);
String category = myCategorizer.getBestOutcome();
				

The following command will train the document categorizer and write the model to en-doccat.bin:

						
$ opennlp DoccatTrainer -model en-doccat.bin -lang en -data en-doccat.train -encoding UTF-8
		 

Additionally it is possible to specify the number of iterations, and the cutoff.

So, naturally you will need some access to many pre-classified events to train your model. The class opennlp.tools.doccat.DocumentSample encapsulates a text document and its classification. DocumentSample has two constructors. Each take the text's category as one argument. The other argument can either be raw text, or an array of tokens. By default, the raw text will be split into tokens by whitespace. So, let's say your training data was contained in a text file, where the format is as described above. Then you might want to write something like this to create a collection of DocumentSamples:

						
DoccatModel model = null;

InputStream dataIn = null;
try {
  dataIn = new FileInputStream("en-sentiment.train");
  ObjectStream<String> lineStream =
		new PlainTextByLineStream(dataIn, "UTF-8");
  ObjectStream<DocumentSample> sampleStream = new DocumentSampleStream(lineStream);

  model = DocumentCategorizerME.train("en", sampleStream);
}
catch (IOException e) {
  // Failed to read or parse training data, training failed
  e.printStackTrace();
}
finally {
  if (dataIn != null) {
    try {
      dataIn.close();
    }
    catch (IOException e) {
      // Not an issue, training already finished.
      // The exception should be logged and investigated
      // if part of a production system.
      e.printStackTrace();
    }
  }
}
	

Now might be a good time to cruise over to Hulu or something, because this could take a while if you've got a large training set. You may see a lot of output as well. Once you're done, you can pretty quickly step to classification directly, but first we'll cover serialization. Feel free to skim.

						
OutputStream modelOut = null;
try {
  modelOut = new BufferedOutputStream(new FileOutputStream(modelFile));
  model.serialize(modelOut);
}
catch (IOException e) {
  // Failed to save model
  e.printStackTrace();
}
finally {
  if (modelOut != null) {
    try {
       modelOut.close();
    }
    catch (IOException e) {
      // Failed to correctly save model.
      // Written model might be invalid.
      e.printStackTrace();
    }
  }
}

The easiest way to try out the POS Tagger is the command line tool. The tool is only intended for demonstration and testing. Download the english maxent pos model and start the POS Tagger Tool with this command:

			
$ opennlp POSTagger en-pos-maxent.bin
		 

The POS Tagger now reads a tokenized sentence per line from stdin. Copy these two sentences to the console:

			
Pierre Vinken , 61 years old , will join the board as a nonexecutive director Nov. 29 .
Mr. Vinken is chairman of Elsevier N.V. , the Dutch publishing group .
		 

the POS Tagger will now echo the sentences with pos tags to the console:

			
Pierre_NNP Vinken_NNP ,_, 61_CD years_NNS old_JJ ,_, will_MD join_VB the_DT board_NN as_IN
    a_DT nonexecutive_JJ director_NN Nov._NNP 29_CD ._.
Mr._NNP Vinken_NNP is_VBZ chairman_NN of_IN Elsevier_NNP N.V._NNP ,_, the_DT Dutch_NNP publishing_VBG group_NN
		 

The tag set used by the english pos model is the Penn Treebank tag set.

The POS Tagger can be embedded into an application via its API. First the pos model must be loaded into memory from disk or an other source. In the sample below its loaded from disk.

				
InputStream modelIn = null;

try {
  modelIn = new FileInputStream("en-pos-maxent.bin");
  POSModel model = new POSModel(modelIn);
}
catch (IOException e) {
  // Model loading failed, handle the error
  e.printStackTrace();
}
finally {
  if (modelIn != null) {
    try {
      modelIn.close();
    }
    catch (IOException e) {
    }
  }
}
			

After the model is loaded the POSTaggerME can be instantiated.

				
POSTaggerME tagger = new POSTaggerME(model);
			

The POS Tagger instance is now ready to tag data. It expects a tokenized sentence as input, which is represented as a String array, each String object in the array is one token.

The following code shows how to determine the most likely pos tag sequence for a sentence.

		  
String sent[] = new String[]{"Most", "large", "cities", "in", "the", "US", "had",
                             "morning", "and", "afternoon", "newspapers", "."};		  
String tags[] = tagger.tag(sent);
			

The tags array contains one part-of-speech tag for each token in the input array. The corresponding tag can be found at the same index as the token has in the input array. The confidence scores for the returned tags can be easily retrieved from a POSTaggerME with the following method call:

		  
double probs[] = tagger.probs();
			

The call to probs is stateful and will always return the probabilities of the last tagged sentence. The probs method should only be called when the tag method was called before, otherwise the behavior is undefined.

Some applications need to retrieve the n-best pos tag sequences and not only the best sequence. The topKSequences method is capable of returning the top sequences. It can be called in a similar way as tag.

		  
Sequence topSequences[] = tagger.topKSequences(sent);
			

Each Sequence object contains one sequence. The sequence can be retrieved via Sequence.getOutcomes() which returns a tags array and Sequence.getProbs() returns the probability array for this sequence.

OpenNLP has a command line tool which is used to train the models available from the model download page on various corpora.

Usage of the tool:

				
$ opennlp POSTaggerTrainer
Usage: opennlp POSTaggerTrainer[.conllx] [-type maxent|perceptron|perceptron_sequence] \
               [-dict dictionaryPath] [-ngram cutoff] [-params paramsFile] [-iterations num] \
               [-cutoff num] -model modelFile -lang language -data sampleData \
               [-encoding charsetName]

Arguments description:
        -type maxent|perceptron|perceptron_sequence
                The type of the token name finder model. One of maxent|perceptron|perceptron_sequence.
        -dict dictionaryPath
                The XML tag dictionary file
        -ngram cutoff
                NGram cutoff. If not specified will not create ngram dictionary.
        -params paramsFile
                training parameters file.
        -iterations num
                number of training iterations, ignored if -params is used.
        -cutoff num
                minimal number of times a feature must be seen, ignored if -params is used.
        -model modelFile
                output model file.
        -lang language
                language which is being processed.
        -data sampleData
                data to be used, usually a file name.
        -encoding charsetName
                encoding for reading and writing text, if absent the system default is used.
			 

The following command illustrates how an english part-of-speech model can be trained:

		  
$ opennlp POSTaggerTrainer -type maxent -model en-pos-maxent.bin \
                           -lang en -data en-pos.train -encoding UTF-8
		    

The Part-of-Speech Tagger training API supports the training of a new pos model. Basically three steps are necessary to train it:

The following code illustrates that:

				
POSModel model = null;

InputStream dataIn = null;
try {
  dataIn = new FileInputStream("en-pos.train");
  ObjectStream<String> lineStream = new PlainTextByLineStream(dataIn, "UTF-8");
  ObjectStream<POSSample> sampleStream = new WordTagSampleStream(lineStream);

  model = POSTaggerME.train("en", sampleStream, TrainingParameters.defaultParams(), null, null);
}
catch (IOException e) {
  // Failed to read or parse training data, training failed
  e.printStackTrace();
}
finally {
  if (dataIn != null) {
    try {
      dataIn.close();
    }
    catch (IOException e) {
      // Not an issue, training already finished.
      // The exception should be logged and investigated
      // if part of a production system.
      e.printStackTrace();
    }
  }
}
	

The above code performs the first two steps, opening the data and training the model. The trained model must still be saved into an OutputStream, in the sample below it is written into a file.

				
OutputStream modelOut = null;
try {
  modelOut = new BufferedOutputStream(new FileOutputStream(modelFile));
  model.serialize(modelOut);
}
catch (IOException e) {
  // Failed to save model
  e.printStackTrace();
}
finally {
  if (modelOut != null) {
  try {
     modelOut.close();
  }
  catch (IOException e) {
    // Failed to correctly save model.
    // Written model might be invalid.
    e.printStackTrace();
  }
}
		

The tag dictionary is a word dictionary which specifies which tags a specific token can have. Using a tag dictionary has two advantages, inappropriate tags can not been assigned to tokens in the dictionary and the beam search algorithm has to consider less possibilities and can search faster.

The dictionary is defined in a xml format and can be created and stored with the POSDictionary class. Please for now checkout the javadoc and source code of that class.

Note: The format should be documented and sample code should show how to use the dictionary. Any contributions are very welcome. If you want to contribute please contact us on the mailing list or comment on the jira issue OPENNLP-287.

There is a command line tool to evaluate a given model on a test data set. The following command shows how the tool can be run:

				
$ opennlp POSTaggerEvaluator -model pt.postagger.bin -lang pt -data pt.postagger.test -encoding utf-8
			 

This will display the resulting accuracy score, e.g.:

				
Loading model ... done
Evaluating ... done

Accuracy: 0.9659110277825124
			 

There is a command line tool to cross validate a test data set. The following command shows how the tool can be run:

                    
$ opennlp POSTaggerCrossValidator -lang pt -data pt.postagger.test -encoding utf-8
                 

This will display the resulting accuracy score, e.g.:

                    
Accuracy: 0.9659110277825124
                 

The easiest way to try out the Chunker is the command line tool. The tool is only intended for demonstration and testing.

Download the english maxent chunker model from the website and start the Chunker Tool with this command:

				
$ opennlp ChunkerME en-chunker.bin
		

The Chunker now reads a pos tagged sentence per line from stdin. Copy these two sentences to the console:

				
Rockwell_NNP International_NNP Corp._NNP 's_POS Tulsa_NNP unit_NN said_VBD it_PRP signed_VBD 
    a_DT tentative_JJ agreement_NN extending_VBG its_PRP$ contract_NN with_IN Boeing_NNP Co._NNP
    to_TO provide_VB structural_JJ parts_NNS for_IN Boeing_NNP 's_POS 747_CD jetliners_NNS ._.
Rockwell_NNP said_VBD the_DT agreement_NN calls_VBZ for_IN it_PRP to_TO supply_VB 200_CD
    additional_JJ so-called_JJ shipsets_NNS for_IN the_DT planes_NNS ._.
		

The Chunker will now echo the sentences grouped tokens to the console:

				
[NP Rockwell_NNP International_NNP Corp._NNP ] [NP 's_POS Tulsa_NNP unit_NN ] [VP said_VBD ]
    [NP it_PRP ] [VP signed_VBD ] [NP a_DT tentative_JJ agreement_NN ] [VP extending_VBG ]
    [NP its_PRP$ contract_NN ] [PP with_IN ] [NP Boeing_NNP Co._NNP ] [VP to_TO provide_VB ]
    [NP structural_JJ parts_NNS ] [PP for_IN ] [NP Boeing_NNP ] [NP 's_POS 747_CD jetliners_NNS ] ._.
[NP Rockwell_NNP ] [VP said_VBD ] [NP the_DT agreement_NN ] [VP calls_VBZ ] [SBAR for_IN ]
    [NP it_PRP ] [VP to_TO supply_VB ] [NP 200_CD additional_JJ so-called_JJ shipsets_NNS ]
    [PP for_IN ] [NP the_DT planes_NNS ] ._.
		

The tag set used by the english pos model is the Penn Treebank tag set.

The Chunker can be embedded into an application via its API. First the chunker model must be loaded into memory from disk or an other source. In the sample below its loaded from disk.

				
InputStream modelIn = null;
ChunkerModel model = null;

try {
  modelIn = new FileInputStream("en-chunker.bin");
  model = new ChunkerModel(modelIn);
} catch (IOException e) {
  // Model loading failed, handle the error
  e.printStackTrace();
} finally {
  if (modelIn != null) {
    try {
      modelIn.close();
    } catch (IOException e) {
    }
  }
}
			

After the model is loaded a Chunker can be instantiated.

				
ChunkerME chunker = new ChunkerME(model);
			

The Chunker instance is now ready to tag data. It expects a tokenized sentence as input, which is represented as a String array, each String object in the array is one token, and the POS tags associated with each token.

The following code shows how to determine the most likely chunk tag sequence for a sentence.

		  
String sent[] = new String[] { "Rockwell", "International", "Corp.", "'s",
    "Tulsa", "unit", "said", "it", "signed", "a", "tentative", "agreement",
    "extending", "its", "contract", "with", "Boeing", "Co.", "to",
    "provide", "structural", "parts", "for", "Boeing", "'s", "747",
    "jetliners", "." };

String pos[] = new String[] { "NNP", "NNP", "NNP", "POS", "NNP", "NN",
    "VBD", "PRP", "VBD", "DT", "JJ", "NN", "VBG", "PRP$", "NN", "IN",
    "NNP", "NNP", "TO", "VB", "JJ", "NNS", "IN", "NNP", "POS", "CD", "NNS",
    "." };

String tag[] = chunker.chunk(sent, pos);
			

The tags array contains one chunk tag for each token in the input array. The corresponding tag can be found at the same index as the token has in the input array. The confidence scores for the returned tags can be easily retrieved from a ChunkerME with the following method call:

		  
double probs[] = chunker.probs();
			

The call to probs is stateful and will always return the probabilities of the last tagged sentence. The probs method should only be called when the tag method was called before, otherwise the behavior is undefined.

Some applications need to retrieve the n-best chunk tag sequences and not only the best sequence. The topKSequences method is capable of returning the top sequences. It can be called in a similar way as chunk.

		  
Sequence topSequences[] = chunk.topKSequences(sent, pos);
			

Each Sequence object contains one sequence. The sequence can be retrieved via Sequence.getOutcomes() which returns a tags array and Sequence.getProbs() returns the probability array for this sequence.

OpenNLP has a command line tool which is used to train the models available from the model download page on various corpora.

Usage of the tool:

				
$ opennlp ChunkerTrainerME
Usage: opennlp ChunkerTrainerME[.ad] [-params paramsFile] [-iterations num] [-cutoff num] \
               -model modelFile -lang language -data sampleData [-encoding charsetName]

Arguments description:
        -params paramsFile
                training parameters file.
        -iterations num
                number of training iterations, ignored if -params is used.
        -cutoff num
                minimal number of times a feature must be seen, ignored if -params is used.
        -model modelFile
                output model file.
        -lang language
                language which is being processed.
        -data sampleData
                data to be used, usually a file name.
        -encoding charsetName
                encoding for reading and writing text, if absent the system default is used.
		

Its now assumed that the english chunker model should be trained from a file called en-chunker.train which is encoded as UTF-8. The following command will train the name finder and write the model to en-chunker.bin:

		
$ opennlp ChunkerTrainerME -model en-chunker.bin -lang en -data en-chunker.train -encoding UTF-8
		

Additionally its possible to specify the number of iterations, the cutoff and to overwrite all types in the training data with a single type.

The Chunker offers an API to train a new chunker model. The following sample code illustrates how to do it:

                    
Charset charset = Charset.forName("UTF-8");
ObjectStream<String> lineStream =
    new PlainTextByLineStream(new FileInputStream("en-chunker.train"),charset);
ObjectStream<ChunkSample> sampleStream = new ChunkSampleStream(lineStream);

ChunkerModel model;

try {
  model = ChunkerME.train("en", sampleStream,
      new DefaultChunkerContextGenerator(), TrainingParameters.defaultParams());
}
finally {
  sampleStream.close();
}

OutputStream modelOut = null;
try {
  modelOut = new BufferedOutputStream(new FileOutputStream(modelFile));
  model.serialize(modelOut);
} finally {
  if (modelOut != null)
     modelOut.close();
}
                

The following command shows how the tool can be run:

				
$ opennlp ChunkerEvaluator
Usage: opennlp ChunkerEvaluator[.ad] -model model [-misclassified true|false] \
               [-detailedF true|false] -lang language -data sampleData [-encoding charsetName]
		

A sample of the command considering you have a data sample named en-chunker.eval and you trained a model called en-chunker.bin:

				
$ opennlp ChunkerEvaluator -model en-chunker.bin -lang en -data en-chunker.eval -encoding UTF-8
		

and here is a sample output:

		
Precision: 0.9255923572240226
Recall: 0.9220610430991112
F-Measure: 0.9238233255623465
		

You can also use the tool to perform 10-fold cross validation of the Chunker. he following command shows how the tool can be run:

				
$ opennlp ChunkerCrossValidator
Usage: opennlp ChunkerCrossValidator[.ad] [-params paramsFile] [-iterations num] [-cutoff num] \
               [-misclassified true|false] [-folds num] [-detailedF true|false] \
               -lang language -data sampleData [-encoding charsetName]

Arguments description:
        -params paramsFile
                training parameters file.
        -iterations num
                number of training iterations, ignored if -params is used.
        -cutoff num
                minimal number of times a feature must be seen, ignored if -params is used.
        -misclassified true|false
                if true will print false negatives and false positives.
        -folds num
                number of folds, default is 10.
        -detailedF true|false
                if true will print detailed FMeasure results.
        -lang language
                language which is being processed.
        -data sampleData
                data to be used, usually a file name.
        -encoding charsetName
                encoding for reading and writing text, if absent the system default is used.
		

It is not necessary to pass a model. The tool will automatically split the data to train and evaluate:

            
$ opennlp ChunkerCrossValidator -lang pt -data en-chunker.cross -encoding UTF-8
		

The easiest way to try out the Parser is the command line tool. The tool is only intended for demonstration and testing. Download the english chunking parser model from the our website and start the Parse Tool with the following command.

				
$ opennlp Parser en-parser.bin en-parser-chunking.bin
		

Loading the big parser model can take several seconds, be patient. Copy this sample sentence to the console.

				
The quick brown fox jumps over the lazy dog .
		

The parser should now print the following to the console.

				
(TOP (NP (NP (DT The) (JJ quick) (JJ brown) (NN fox) (NNS jumps)) (PP (IN over) (NP (DT the)
    (JJ lazy) (NN dog))) (. .)))
		

With the following command the input can be read from a file and be written to an output file.

				
$ opennlp Parser en-parser.bin en-parser-chunking.bin < article-tokenized.txt > article-parsed.txt.
		

The article-tokenized.txt file must contain one sentence per line which is tokenized with the english tokenizer model from our website. See the Tokenizer documentation for further details.

The Parser can be easily integrated into an application via its API. To instantiate a Parser the parser model must be loaded first.

				
InputStream modelIn = new FileInputStream("en-parser-chunking.bin");
try {
  ParserModel model = new ParserModel(modelIn);
}
catch (IOException e) {
  e.printStackTrace();
}
finally {
  if (modelIn != null) {
    try {
      modelIn.close();
    }
    catch (IOException e) {
    }
  }
}
		

Unlike the other components to instantiate the Parser a factory method should be used instead of creating the Parser via the new operator. The parser model is either trained for the chunking parser or the tree insert parser the parser implementation must be chosen correctly. The factory method will read a type parameter from the model and create an instance of the corresponding parser implementation.

				
Parser parser = ParserFactory.create(model);
		

Right now the tree insert parser is still experimental and there is no pre-trained model for it. The parser expect a whitespace tokenized sentence. A utility method from the command line tool can parse the sentence String. The following code shows how the parser can be called.

				
String sentence = "The quick brown fox jumps over the lazy dog .";
Parse topParses[] = ParserTool.parseLine(sentence, parser, 1);
		

The topParses array only contains one parse because the number of parses is set to 1. The Parse object contains the parse tree. To display the parse tree call the show method. It either prints the parse to the console or into a provided StringBuffer. Similar to Exception.printStackTrace.

TODO: Extend this section with more information about the Parse object.

OpenNLP has a command line tool which is used to train the models available from the model download page on various corpora. The data must be converted to the OpenNLP parser training format, which is shortly explained above. To train the parser a head rules file is also needed. (TODO: Add documentation about the head rules file) Usage of the tool:

				
$ opennlp ParserTrainer
Usage: opennlp ParserTrainer -headRules headRulesFile [-parserType CHUNKING|TREEINSERT] \
                             [-params paramsFile] [-iterations num] [-cutoff num] \
                             -model modelFile -lang language -data sampleData \
                             [-encoding charsetName]

Arguments description:
	-headRules headRulesFile
		head rules file.
	-parserType CHUNKING|TREEINSERT
		one of CHUNKING or TREEINSERT, default is CHUNKING.
	-params paramsFile
		training parameters file.
	-iterations num
		number of training iterations, ignored if -params is used.
	-cutoff num
		minimal number of times a feature must be seen, ignored if -params is used.
	-model modelFile
		output model file.
	-format formatName
		data format, might have its own parameters.
	-encoding charsetName
		encoding for reading and writing text, if absent the system default is used.
	-lang language
		language which is being processed.
	-data sampleData
		data to be used, usually a file name.
	-encoding charsetName
		encoding for reading and writing text, if absent the system default is used.
        

The model on the website was trained with the following command:

		
$ opennlp ParserTrainer -model en-parser-chunking.bin -parserType CHUNKING \
                        -head-rules head_rules \
                        -lang en -data train.all -encoding ISO-8859-1
    
		

Its also possible to specify the cutoff and the number of iterations, these parameters are used for all trained models. The -parserType parameter is an optional parameter, to use the tree insertion parser, specify TREEINSERT as type. The TaggerModelReplacer tool replaces the tagger model inside the parser model with a new one.

Note: The original parser model will be overwritten with the new parser model which contains the replaced tagger model.

		
$ opennlp TaggerModelReplacer en-parser-chunking.bin en-pos-maxent.bin
		

Additionally there are tools to just retrain the build or the check model.

TODO: Write documentation about the parser training api. Any contributions are very welcome. If you want to contribute please contact us on the mailing list or comment on the jira issue OPENNLP-219.

The shared task of CoNLL-2000 is Chunking.

			
$ opennlp ChunkerTrainerME -model en-chunker.bin -iterations 500 \
                           -lang en -data train.txt -encoding UTF-8
		

			
Indexing events using cutoff of 5

	Computing event counts...  done. 211727 events
	Indexing...  done.
Sorting and merging events... done. Reduced 211727 events to 197252.
Done indexing.
Incorporating indexed data for training...  
done.
	Number of Event Tokens: 197252
	    Number of Outcomes: 22
	  Number of Predicates: 107838
...done.
Computing model parameters...
Performing 500 iterations.
  1:  .. loglikelihood=-654457.1455212828	0.2601510435608118
  2:  .. loglikelihood=-239513.5583724216	0.9260037690044255
  3:  .. loglikelihood=-141313.1386347238	0.9443387003074715
  4:  .. loglikelihood=-101083.50853437989	0.954375209585929
... cut lots of iterations ...
498:  .. loglikelihood=-1710.8874647317095	0.9995040783650645
499:  .. loglikelihood=-1708.0908900815848	0.9995040783650645
500:  .. loglikelihood=-1705.3045902366732	0.9995040783650645
Writing chunker model ... done (4.019s)

Wrote chunker model to path: .\en-chunker.bin
		

			
$ opennlp ChunkerEvaluator -model en-chunker.bin -lang en -encoding utf8 -data test.txt
		

			
Loading Chunker model ... done (0,665s)
current: 85,8 sent/s avg: 85,8 sent/s total: 86 sent
current: 88,1 sent/s avg: 87,0 sent/s total: 174 sent
current: 156,2 sent/s avg: 110,0 sent/s total: 330 sent
current: 192,2 sent/s avg: 130,5 sent/s total: 522 sent
current: 167,2 sent/s avg: 137,8 sent/s total: 689 sent
current: 179,2 sent/s avg: 144,6 sent/s total: 868 sent
current: 183,2 sent/s avg: 150,3 sent/s total: 1052 sent
current: 183,2 sent/s avg: 154,4 sent/s total: 1235 sent
current: 169,2 sent/s avg: 156,0 sent/s total: 1404 sent
current: 178,2 sent/s avg: 158,2 sent/s total: 1582 sent
current: 172,2 sent/s avg: 159,4 sent/s total: 1754 sent
current: 177,2 sent/s avg: 160,9 sent/s total: 1931 sent


Average: 161,6 sent/s 
Total: 2013 sent
Runtime: 12.457s

Precision: 0.9244354736974896
Recall: 0.9216837162502096
F-Measure: 0.9230575441395671
		

The shared task of CoNLL-2002 is language independent named entity recognition for Spanish and Dutch.

			
$ opennlp TokenNameFinderConverter conll02 -data esp.train -lang es -types per > es_corpus_train_persons.txt
		

			
$ opennlp TokenNameFinderConverter conll02 -data esp.testa -lang es -types per > corpus_testa.txt
$ opennlp TokenNameFinderConverter conll02 -data esp.testb -lang es -types per > corpus_testb.txt
		

			
\bin\opennlp TokenNameFinderTrainer -lang es -encoding u
tf8 -iterations 500 -data es_corpus_train_persons.txt -model es_ner_person.bin


Indexing events using cutoff of 5

        Computing event counts...  done. 264715 events
        Indexing...  done.
Sorting and merging events... done. Reduced 264715 events to 222660.
Done indexing.
Incorporating indexed data for training...
done.
        Number of Event Tokens: 222660
           Number of Outcomes: 3
          Number of Predicates: 71514
...done.
Computing model parameters ...
Performing 500 iterations.
  1:  ... loglikelihood=-290819.1519958615      0.9689326256540053
  2:  ... loglikelihood=-37097.17676455632      0.9689326256540053
  3:  ... loglikelihood=-22910.372489660916     0.9706476776911017
  4:  ... loglikelihood=-17091.547325669497     0.9777874317662392
  5:  ... loglikelihood=-13797.620926769372     0.9833821279489262
  6:  ... loglikelihood=-11715.806710780415     0.9867140131839903
  7:  ... loglikelihood=-10289.222078246517     0.9886859452618855
  8:  ... loglikelihood=-9249.208318314624      0.9902310031543358
  9:  ... loglikelihood=-8454.169590899777      0.9913227433277298
 10:  ... loglikelihood=-7823.742997451327      0.9921953799369133
 11:  ... loglikelihood=-7309.375882641964      0.9928224694482746
 12:  ... loglikelihood=-6880.131972149693      0.9932946754056249
 13:  ... loglikelihood=-6515.3828767792365     0.993638441342576
 14:  ... loglikelihood=-6200.82723154046       0.9939595413935742
 15:  ... loglikelihood=-5926.213730444915      0.994269308501596
 16:  ... loglikelihood=-5683.9821840753275     0.9945299661900534
 17:  ... loglikelihood=-5468.4211798176075     0.9948246227074401
 18:  ... loglikelihood=-5275.127017232056      0.9950286156810154

... cut lots of iterations ...

491:  ... loglikelihood=-1174.8485558758211     0.998983812779782
492:  ... loglikelihood=-1173.9971776942477     0.998983812779782
493:  ... loglikelihood=-1173.1482915871768     0.998983812779782
494:  ... loglikelihood=-1172.3018855781158     0.998983812779782
495:  ... loglikelihood=-1171.457947774544      0.998983812779782
496:  ... loglikelihood=-1170.6164663670502     0.998983812779782
497:  ... loglikelihood=-1169.7774296286693     0.998983812779782
498:  ... loglikelihood=-1168.94082591387       0.998983812779782
499:  ... loglikelihood=-1168.1066436580463     0.9989875904274408
500:  ... loglikelihood=-1167.2748713765225     0.9989875904274408
Writing name finder model ... done (2,168s)

Wrote name finder model to
path: .\es_ner_person.bin
		

The shared task of CoNLL-2003 is language independent named entity recognition for English and German.

			
$ opennlp TokenNameFinderConverter conll03 -lang en -types per -data eng.train > corpus_train.txt
		

			
$ opennlp TokenNameFinderConverter conll03 -lang en -types per -data eng.testa > corpus_testa.txt
$ opennlp TokenNameFinderConverter conll03 -lang en -types per -data eng.testb > corpus_testb.txt
		

                
$ opennlp TokenNameFinderTrainer.conll03 -model en_ner_person.bin -iterations 500 \
                                 -lang en -types per -data eng.train -encoding utf8
                

                
$ opennlp TokenNameFinderTrainer -model en_ner_person.bin -iterations 500 \
                                 -lang en -data corpus_train.txt -encoding utf8
		        

			    
Indexing events using cutoff of 5

	Computing event counts...  done. 203621 events
	Indexing...  done.
Sorting and merging events... done. Reduced 203621 events to 179409.
Done indexing.
Incorporating indexed data for training...  
done.
	Number of Event Tokens: 179409
	    Number of Outcomes: 3
	  Number of Predicates: 58814
...done.
Computing model parameters...
Performing 500 iterations.
  1:  .. loglikelihood=-223700.5328318588	0.9453494482396216
  2:  .. loglikelihood=-40525.939777363084	0.9467933071736215
  3:  .. loglikelihood=-24893.98837874921	0.9598518816821447
  4:  .. loglikelihood=-18420.3379471033	0.9712996203731442
... cut lots of iterations ...
498:  .. loglikelihood=-952.8501399442295	0.9988950059178572
499:  .. loglikelihood=-952.0600155746948	0.9988950059178572
500:  .. loglikelihood=-951.2722802086295	0.9988950059178572
Writing name finder model ... done (1.638s)

Wrote name finder model to
path: .\en_ner_person.bin
		        

                
$ opennlp TokenNameFinderEvaluator.conll03 -model en_ner_person.bin \
                                   -lang en -types per -data eng.testa -encoding utf8
                

			
$ opennlp TokenNameFinderEvaluator -model en_ner_person.bin -lang en -data corpus_testa.txt \
                                   -encoding utf8
		        

			
Loading Token Name Finder model ... done (0.359s)
current: 190.2 sent/s avg: 190.2 sent/s total: 199 sent
current: 648.3 sent/s avg: 415.9 sent/s total: 850 sent
current: 530.1 sent/s avg: 453.6 sent/s total: 1380 sent
current: 793.8 sent/s avg: 539.0 sent/s total: 2178 sent
current: 705.4 sent/s avg: 571.9 sent/s total: 2882 sent


Average: 569.4 sent/s
Total: 3251 sent
Runtime: 5.71s

Precision: 0.9366247297154147
Recall: 0.739956568946797
F-Measure: 0.8267557582133971
		

The Corpus can be downloaded from here: http://www.linguateca.pt/floresta/corpus.html

The Name Finder models were trained using the Amazonia corpus: amazonia.ad. The Chunker models were trained using the Bosque_CF_8.0.ad.

To extract NameFinder training data from Amazonia corpus:

			        
$ opennlp TokenNameFinderConverter ad -lang pt -encoding ISO-8859-1 -data amazonia.ad > corpus.txt
			        

To extract Chunker training data from Bosque_CF_8.0.ad corpus:

			    
$ opennlp ChunkerConverter ad -lang pt -data Bosque_CF_8.0.ad.txt -encoding ISO-8859-1 > bosque-chunk
    			

To perform the evaluation the corpus was split into a training and a test part.

			        
$ sed '1,55172d' corpus.txt > corpus_train.txt
$ sed '55172,100000000d' corpus.txt > corpus_test.txt
        			

        			
$ opennlp TokenNameFinderTrainer -model pt-ner.bin -cutoff 20 -lang PT -data corpus_train.txt -encoding UTF-8
...
$ opennlp TokenNameFinderEvaluator -model pt-ner.bin -lang PT -data corpus_train.txt -encoding UTF-8

Precision: 0.8005071889818507
Recall: 0.7450581122145297
F-Measure: 0.7717879983140168
			

We have tried to make the opennlp.maxent implementation easy to use. To create a model, one needs (of course) the training data, and then implementations of two interfaces in the opennlp.maxent package, EventStream and ContextGenerator. These have fairly simple specifications, and example implementations can be found in the OpenNLP Tools preprocessing components.

We have also set in place some interfaces and code to make it easier to automate the training and evaluation process (the Evalable interface and the TrainEval class). It is not necessary to use this functionality, but if you do you'll find it much easier to see how well your models are doing. The opennlp.grok.preprocess.namefind package is an example of a maximum entropy component which uses this functionality.

We have managed to use several techniques to reduce the size of the models when writing them to disk, which also means that reading in a model for use is much quicker than with less compact encodings of the model. This was especially important to us since we use many maxent models in the Grok library, and we wanted the start up time and the physical size of the library to be as minimal as possible. As of version 1.2.0, maxent has an io package which greatly simplifies the process of loading and saving models in different formats.