Package org.coordinatekit.crf.mallet.train

@NullMarked package org.coordinatekit.crf.mallet.train

MALLET-based CRF training implementation.

This package provides the MALLET-based implementation for training CRF models using the MALLET (MAchine Learning for LanguagE Toolkit) library.

Example Usage

 
 // Configure model checkpoint output
 ModelOutputConfiguration modelOutput = ModelOutputConfiguration.builder()
         .outputDirectory(Path.of("checkpoints"))
         .iterationInterval(50)
         .build();

 // Configure CoNLL evaluation output
 ConllOutputConfiguration conllOutput = ConllOutputConfiguration.builder()
         .outputDirectory(Path.of("evaluations"))
         .iterationInterval(50)
         .build();

 // Configure training parameters
 MalletCrfTrainerConfiguration config = MalletCrfTrainerConfiguration.builder()
         .iterations(500)
         .gaussianVariance(10.0)
         .threads(8)
         .trainingFraction(0.8)
         .modelOutputConfiguration(modelOutput)
         .conllOutputConfiguration(conllOutput)
         .build();

 // Create the trainer
 MalletCrfTrainer<String, MyTag> trainer = new MalletCrfTrainer<>(
         featureExtractor,
         tagProvider,
         trainingDataSequencer,
         config);

 // Train and save the model
 trainer.train(Path.of("training-data.xml"), Path.of("model.ser"));
 
 

Configuration Options

Key parameters in MalletCrfTrainerConfiguration:

• iterations - Maximum training iterations (default: 500)
• gaussianVariance - L2 regularization strength; higher values = less regularization (default: 10.0)
• threads - Number of threads for parallel training (default: 6)
• trainingFraction - Fraction of data for training vs. testing (default: 0.5)
• weightsType - Memory/speed trade-off for weight storage (default: SOME_DENSE)

Parameters in ModelOutputConfiguration for saving model checkpoints:

• outputDirectory - Directory for checkpoint files
• iterationInterval - Save model every N iterations (default: 10)
• filePrefix - Prefix for checkpoint filenames (default: "model_")
• fileSuffix - Suffix/extension for checkpoint files (default: ".ser")

Parameters in ConllOutputConfiguration for CoNLL-format evaluation output:

• outputDirectory - Directory for CoNLL output files
• iterationInterval - Write predictions every N iterations (default: 10)
• filePrefix - Prefix for output filenames (default: "output_iter")
• fileSuffix - Suffix/extension for output files (default: ".conll")

See Also:

• MalletCrfTrainer
• MalletCrfTrainerConfiguration
• CrfTrainer

Class	Description
CompositeTestAccuracyEvaluator	A composite evaluator that logs iteration statistics along with both instance and token accuracy metrics on test data.
ConllOutputConfiguration	Configuration settings for ConllOutputEvaluator.
ConllOutputConfiguration.Builder	Builder for constructing ConllOutputConfiguration instances.
ConllOutputEvaluator	An evaluator that outputs predicted and actual labels in CoNLL format.
MalletCrfTrainer<F,T extends Comparable<T>>	A CRF trainer implementation using the MALLET (MAchine Learning for LanguagE Toolkit) library.
MalletCrfTrainerConfiguration	Configuration settings for MalletCrfTrainer.
MalletCrfTrainerConfiguration.Builder	Builder for constructing MalletCrfTrainerConfiguration instances.
ModelOutputConfiguration	Configuration settings for ModelOutputEvaluator.
ModelOutputConfiguration.Builder	Builder for constructing ModelOutputConfiguration instances.
ModelOutputEvaluator	An evaluator that writes the current transducer (model) to file using Java serialization.
SimpleTrainingTestSplit	A container for training and test data splits.
TrainingTestSplit	A container for training and test data splits.
WeightsType	Specifies the weight storage strategy for CRF training.