cc.mallet.fst.semi_supervised.tui
Class SimpleTaggerWithConstraints

java.lang.Object
  extended by cc.mallet.fst.semi_supervised.tui.SimpleTaggerWithConstraints

public class SimpleTaggerWithConstraints
extends java.lang.Object

Version of SimpleTagger that trains CRFs with expectation constraints rather than labeled data. This class's main method trains, tests, or runs a generic CRF-based sequence tagger.

Training and test files consist of blocks of lines, one block for each instance, separated by blank lines. Each block of lines should have the first form specified for the input of SimpleTagger.SimpleTaggerSentence2FeatureVectorSequence. A variety of command line options control the operation of the main program, as described in the comments for main.

Version:
1.0
Author:
Gregory Druck gdruck@cs.umass.edu

Method Summary
static Sequence[] apply(Transducer model, Sequence input, int k)
          Apply a transducer to an input sequence to produce the k highest-scoring output sequences.
static CRF getCRF(InstanceList training, int[] orders, java.lang.String defaultLabel, java.lang.String forbidden, java.lang.String allowed, boolean connected)
           
static void main(java.lang.String[] args)
          Command-line wrapper to train, test, or run a generic CRF-based tagger.
static void test(TransducerTrainer tt, TransducerEvaluator eval, InstanceList testing)
          Test a transducer on the given test data, evaluating accuracy with the given evaluator
static CRF trainGE(InstanceList training, InstanceList testing, java.util.ArrayList<GEConstraint> constraints, CRF crf, TransducerEvaluator eval, int iterations, double var, int resets)
          Create and train a CRF model from the given training data, optionally testing it on the given test data.
static CRF trainPR(InstanceList training, InstanceList testing, java.util.ArrayList<PRConstraint> constraints, CRF crf, TransducerEvaluator eval, int iterations, double var)
          Create and train a CRF model from the given training data, optionally testing it on the given test data.
 
Methods inherited from class java.lang.Object
clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait
 

Method Detail

trainGE

public static CRF trainGE(InstanceList training,
                          InstanceList testing,
                          java.util.ArrayList<GEConstraint> constraints,
                          CRF crf,
                          TransducerEvaluator eval,
                          int iterations,
                          double var,
                          int resets)
Create and train a CRF model from the given training data, optionally testing it on the given test data.

Parameters:
training - training data
testing - test data (possibly null)
constraints - constraints
crf - model
eval - accuracy evaluator (possibly null)
iterations - number of training iterations
var - Gaussian prior variance
resets - Number of resets.
Returns:
the trained model

trainPR

public static CRF trainPR(InstanceList training,
                          InstanceList testing,
                          java.util.ArrayList<PRConstraint> constraints,
                          CRF crf,
                          TransducerEvaluator eval,
                          int iterations,
                          double var)
Create and train a CRF model from the given training data, optionally testing it on the given test data.

Parameters:
training - training data
testing - test data (possibly null)
constraints - constraints
crf - model
eval - accuracy evaluator (possibly null)
iterations - number of training iterations
var - Gaussian prior variance
Returns:
the trained model

getCRF

public static CRF getCRF(InstanceList training,
                         int[] orders,
                         java.lang.String defaultLabel,
                         java.lang.String forbidden,
                         java.lang.String allowed,
                         boolean connected)

test

public static void test(TransducerTrainer tt,
                        TransducerEvaluator eval,
                        InstanceList testing)
Test a transducer on the given test data, evaluating accuracy with the given evaluator

Parameters:
model - a Transducer
eval - accuracy evaluator
testing - test data

apply

public static Sequence[] apply(Transducer model,
                               Sequence input,
                               int k)
Apply a transducer to an input sequence to produce the k highest-scoring output sequences.

Parameters:
model - the Transducer
input - the input sequence
k - the number of answers to return
Returns:
array of the k highest-scoring output sequences

main

public static void main(java.lang.String[] args)
                 throws java.lang.Exception
Command-line wrapper to train, test, or run a generic CRF-based tagger.

Parameters:
args - the command line arguments. Options (shell and Java quoting should be added as needed):
--help boolean
Print this command line option usage information. Give true for longer documentation. Default is false.
--prefix-code Java-code
Java code you want run before any other interpreted code. Note that the text is interpreted without modification, so unlike some other Java code options, you need to include any necessary 'new's. Default is null.
--gaussian-variance positive-number
The Gaussian prior variance used for training. Default is 10.0.
--train boolean
Whether to train. Default is false.
--iterations positive-integer
Number of training iterations. Default is 500.
--test lab or seg=start-1.continue-1,...,start-n.continue-n
Test measuring labeling or segmentation (start-i, continue-i) accuracy. Default is no testing.
--training-proportion number-between-0-and-1
Fraction of data to use for training in a random split. Default is 0.5.
--model-file filename
The filename for reading (train/run) or saving (train) the model. Default is null.
--random-seed integer
The random seed for randomly selecting a proportion of the instance list for training Default is 0.
--orders comma-separated-integers
List of label Markov orders (main and backoff) Default is 1.
--forbidden regular-expression
If label-1,label-2 matches the expression, the corresponding transition is forbidden. Default is \\s (nothing forbidden).
--allowed regular-expression
If label-1,label-2 does not match the expression, the corresponding expression is forbidden. Default is .* (everything allowed).
--default-label string
Label for initial context and uninteresting tokens. Default is O.
--viterbi-output boolean
Print Viterbi periodically during training. Default is false.
--fully-connected boolean
Include all allowed transitions, even those not in training data. Default is true.
--weights sparse|some-dense|dense
Create sparse, some dense (using a heuristic), or dense features on transitions. Default is some-dense.
--n-best positive-integer
Number of answers to output when applying model. Default is 1.
--include-input boolean
Whether to include input features when printing decoding output. Default is false.
--threads positive-integer
Number of threads for CRF training. Default is 1.
Remaining arguments:
  • training-data-file if training
  • training-and-test-data-file, if training and testing with random split
  • training-data-file test-data-file if training and testing from separate files
  • test-data-file if testing
  • input-data-file if applying to new data (unlabeled)
Throws:
java.lang.Exception - if an error occurs