/** This is (mostly) copied from CRF4.java */
public boolean[][] labelConnectionsIn(
Alphabet outputAlphabet, InstanceList trainingSet, String start) {
int numLabels = outputAlphabet.size();
boolean[][] connections = new boolean[numLabels][numLabels];
for (int i = 0; i < trainingSet.size(); i++) {
Instance instance = trainingSet.getInstance(i);
FeatureSequence output = (FeatureSequence) instance.getTarget();
for (int j = 1; j < output.size(); j++) {
int sourceIndex = outputAlphabet.lookupIndex(output.get(j - 1));
int destIndex = outputAlphabet.lookupIndex(output.get(j));
assert (sourceIndex >= 0 && destIndex >= 0);
connections[sourceIndex][destIndex] = true;
}
}
// Handle start state
if (start != null) {
int startIndex = outputAlphabet.lookupIndex(start);
for (int j = 0; j < outputAlphabet.size(); j++) {
connections[startIndex][j] = true;
}
}
return connections;
}
public TestCRFPipe(String trainingFilename) throws IOException {
ArrayList<Pipe> pipes = new ArrayList<Pipe>();
PrintWriter out = new PrintWriter("test.out");
int[][] conjunctions = new int[3][];
conjunctions[0] = new int[] {-1};
conjunctions[1] = new int[] {1};
conjunctions[2] = new int[] {-2, -1};
pipes.add(new SimpleTaggerSentence2TokenSequence());
// pipes.add(new FeaturesInWindow("PREV-", -1, 1));
// pipes.add(new FeaturesInWindow("NEXT-", 1, 2));
pipes.add(new OffsetConjunctions(conjunctions));
pipes.add(new TokenTextCharSuffix("C1=", 1));
pipes.add(new TokenTextCharSuffix("C2=", 2));
pipes.add(new TokenTextCharSuffix("C3=", 3));
pipes.add(new RegexMatches("CAPITALIZED", Pattern.compile("^\\p{Lu}.*")));
pipes.add(new RegexMatches("STARTSNUMBER", Pattern.compile("^[0-9].*")));
pipes.add(new RegexMatches("HYPHENATED", Pattern.compile(".*\\-.*")));
pipes.add(new RegexMatches("DOLLARSIGN", Pattern.compile("\\$.*")));
pipes.add(new TokenFirstPosition("FIRSTTOKEN"));
pipes.add(new TokenSequence2FeatureVectorSequence());
pipes.add(new SequencePrintingPipe(out));
Pipe pipe = new SerialPipes(pipes);
InstanceList trainingInstances = new InstanceList(pipe);
trainingInstances.addThruPipe(
new LineGroupIterator(
new BufferedReader(
new InputStreamReader(new GZIPInputStream(new FileInputStream(trainingFilename)))),
Pattern.compile("^\\s*$"),
true));
out.close();
}
public static CRF4 createCRF(File trainingFile, CRFInfo crfInfo) throws FileNotFoundException {
Reader trainingFileReader = new FileReader(trainingFile);
// Create a pipe that we can use to convert the training
// file to a feature vector sequence.
Pipe p = new SimpleTagger.SimpleTaggerSentence2FeatureVectorSequence();
// The training file does contain tags (aka targets)
p.setTargetProcessing(true);
// Register the default tag with the pipe, by looking it up
// in the targetAlphabet before we look up any other tag.
p.getTargetAlphabet().lookupIndex(crfInfo.defaultLabel);
// Create a new instancelist to hold the training data.
InstanceList trainingData = new InstanceList(p);
// Read in the training data.
trainingData.add(new LineGroupIterator(trainingFileReader, Pattern.compile("^\\s*$"), true));
// Create the CRF model.
CRF4 crf = new CRF4(p, null);
// Set various config options
crf.setGaussianPriorVariance(crfInfo.gaussianVariance);
crf.setTransductionType(crfInfo.transductionType);
// Set up the model's states.
if (crfInfo.stateInfoList != null) {
Iterator stateIter = crfInfo.stateInfoList.iterator();
while (stateIter.hasNext()) {
CRFInfo.StateInfo state = (CRFInfo.StateInfo) stateIter.next();
crf.addState(
state.name,
state.initialCost,
state.finalCost,
state.destinationNames,
state.labelNames,
state.weightNames);
}
} else if (crfInfo.stateStructure == CRFInfo.FULLY_CONNECTED_STRUCTURE)
crf.addStatesForLabelsConnectedAsIn(trainingData);
else if (crfInfo.stateStructure == CRFInfo.HALF_CONNECTED_STRUCTURE)
crf.addStatesForHalfLabelsConnectedAsIn(trainingData);
else if (crfInfo.stateStructure == CRFInfo.THREE_QUARTERS_CONNECTED_STRUCTURE)
crf.addStatesForThreeQuarterLabelsConnectedAsIn(trainingData);
else if (crfInfo.stateStructure == CRFInfo.BILABELS_STRUCTURE)
crf.addStatesForBiLabelsConnectedAsIn(trainingData);
else throw new RuntimeException("Unexpected state structure " + crfInfo.stateStructure);
// Set up the weight groups.
if (crfInfo.weightGroupInfoList != null) {
Iterator wgIter = crfInfo.weightGroupInfoList.iterator();
while (wgIter.hasNext()) {
CRFInfo.WeightGroupInfo wg = (CRFInfo.WeightGroupInfo) wgIter.next();
FeatureSelection fs =
FeatureSelection.createFromRegex(
crf.getInputAlphabet(), Pattern.compile(wg.featureSelectionRegex));
crf.setFeatureSelection(crf.getWeightsIndex(wg.name), fs);
}
}
// Train the CRF.
crf.train(trainingData, null, null, null, crfInfo.maxIterations);
return crf;
}