/**
* Parses the development set with the specified grammar and FOM. Returns the accuracy (F1) of the
* resulting parses. Uses the specified <code>beamWidth</code> for all cells spanning >= 2 words.
* For lexical cells, uses <code>beamWidth</code> x 3 (since the FOM does not prioritize entries
* in lexical cells), and allocates <code>beamWidth</code> entries in lexical cells for unary
* productions.
*
* @param sparseMatrixGrammar
* @param posFom
* @param cycle
* @return Accuracy (F1) and speed (w/s)
*/
protected float[] parseDevSet(
final LeftCscSparseMatrixGrammar sparseMatrixGrammar,
final BoundaryPosModel posFom,
final int cycle) {
// Initialize the parser
final ParserDriver opts = new ParserDriver();
opts.researchParserType = ResearchParserType.CartesianProductHashMl;
opts.cellSelectorModel = ccModel;
opts.fomModel = posFom;
// Set beam-width configuration properties
GlobalConfigProperties.singleton()
.setProperty(Parser.PROPERTY_MAX_BEAM_WIDTH, Integer.toString(beamWidth));
GlobalConfigProperties.singleton()
.setProperty(Parser.PROPERTY_LEXICAL_ROW_BEAM_WIDTH, Integer.toString(beamWidth * 3));
GlobalConfigProperties.singleton()
.setProperty(Parser.PROPERTY_LEXICAL_ROW_UNARIES, Integer.toString(beamWidth));
// Parse the dev-set
final CartesianProductHashSpmlParser parser =
new CartesianProductHashSpmlParser(opts, sparseMatrixGrammar);
final long t0 = System.currentTimeMillis();
int words = 0;
final BracketEvaluator evaluator = new BracketEvaluator();
for (final String inputTree : developmentSet) {
parser.parseSentence(inputTree).evaluate(evaluator);
final NaryTree<String> naryTree = NaryTree.read(inputTree, String.class);
words += naryTree.leaves();
}
final long t1 = System.currentTimeMillis();
return new float[] {(float) evaluator.accumulatedResult().f1(), words * 1000f / (t1 - t0)};
}
public GrammarParallelCscSpmvParser(
final ParserDriver opts, final LeftCscSparseMatrixGrammar grammar) {
super(opts, grammar);
final ConfigProperties props = GlobalConfigProperties.singleton();
// Split the binary grammar rules into segments of roughly equal size
final int requestedThreads = props.getIntProperty(ParserDriver.OPT_GRAMMAR_THREAD_COUNT);
final int[] segments = new int[requestedThreads + 1];
final int segmentSize = grammar.cscBinaryRowIndices.length / requestedThreads + 1;
segments[0] = 0;
int i = 1;
// Examine each populated column
for (int j = 1; j < grammar.cscBinaryPopulatedColumns.length - 1; j++) {
if (grammar.cscBinaryPopulatedColumnOffsets[j]
- grammar.cscBinaryPopulatedColumnOffsets[segments[i - 1]]
>= segmentSize) {
segments[i++] = j;
}
}
segments[i] = grammar.cscBinaryPopulatedColumnOffsets.length - 1;
this.grammarThreads = i;
this.cpvSegments = grammarThreads * 2;
GlobalConfigProperties.singleton()
.setProperty(
ParserDriver.RUNTIME_CONFIGURED_THREAD_COUNT,
Integer.toString(
props.getIntProperty(ParserDriver.OPT_CELL_THREAD_COUNT, 1) * grammarThreads));
this.binaryRowSegments = new int[i + 1];
System.arraycopy(segments, 0, binaryRowSegments, 0, binaryRowSegments.length);
if (BaseLogger.singleton().isLoggable(Level.FINE)) {
final StringBuilder sb = new StringBuilder();
for (int j = 1; j < binaryRowSegments.length; j++) {
sb.append(
(grammar.cscBinaryPopulatedColumnOffsets[binaryRowSegments[j]]
- grammar.cscBinaryPopulatedColumnOffsets[binaryRowSegments[j - 1]])
+ " ");
}
BaseLogger.singleton().fine("INFO: CSC Binary Grammar segments of length: " + sb.toString());
}
// Temporary cell storage for each grammar-level thread
this.threadLocalTemporaryCellArrays =
new ThreadLocal<PackedArrayChart.TemporaryChartCell[]>() {
@Override
protected PackedArrayChart.TemporaryChartCell[] initialValue() {
final PackedArrayChart.TemporaryChartCell[] tcs =
new PackedArrayChart.TemporaryChartCell[grammarThreads];
for (int j = 0; j < grammarThreads; j++) {
tcs[j] = new PackedArrayChart.TemporaryChartCell(grammar, false);
}
return tcs;
}
};
}
/**
* Evaluates speed and accuracy with the baseline (fully-split) grammar
*
* @param grammar
* @param lexicon
* @param cycle
*/
@Override
public void initMergeCycle(final Grammar grammar, final Lexicon lexicon, final int cycle) {
this.splitGrammar = grammar;
this.splitLexicon = lexicon;
final String[] split =
GlobalConfigProperties.singleton()
.getProperty(PROPERTY_BEAM_WIDTHS, DEFAULT_BEAM_WIDTHS)
.split(",");
this.beamWidth = Integer.parseInt(split[cycle - 1]);
// Convert the grammar to BUBS sparse-matrix format and train a Boundary POS FOM
BaseLogger.singleton()
.info("Constrained parsing the training-set and training a prioritization model");
final LeftCscSparseMatrixGrammar sparseMatrixGrammar =
convertGrammarToSparseMatrix(splitGrammar, splitLexicon);
final BoundaryPosModel posFom = trainPosFom(sparseMatrixGrammar);
// Record accuracy with the full split grammar (so we can later compare with MergeCandidates)
BaseLogger.singleton().info("Parsing the dev-set with the fully-split grammar");
final float[] parseResult = parseDevSet(sparseMatrixGrammar, posFom, beamWidth);
splitF1 = parseResult[0];
splitSpeed = parseResult[1];
BaseLogger.singleton()
.info(String.format("F1 = %.3f Speed = %.3f w/s", splitF1 * 100, splitSpeed));
}
/**
* Performs pruned inference on a development set. The 'normal' implementation ( {@link
* DiscriminativeMergeObjectiveFunction}) evaluates each merge candidate separately; {@link
* SamplingMergeObjective} shares much of the infrastructure, but evaluate a set of merge candidates
* in combination.
*
* @author Aaron Dunlop
*/
public abstract class InferenceInformedMergeObjectiveFunction extends MergeObjectiveFunction {
protected float splitF1;
protected float splitSpeed;
protected Grammar splitGrammar;
protected Lexicon splitLexicon;
private float minRuleProbability;
private List<String> trainingCorpus;
private List<String> developmentSet;
private CompleteClosureModel ccModel;
protected int beamWidth;
private static final String PROPERTY_BEAM_WIDTHS = "beamWidths";
private static final String DEFAULT_BEAM_WIDTHS = "15,15,20,20,30,30";
private static final String PROPERTY_PARSE_FRACTION = "discParseFraction";
private static final float DEFAULT_PARSE_FRACTION = .5f;
/**
* Fraction of merge candidates to parse - if x < 1, we estimate likelihood loss (as in the
* Likelihood {@link MergeRanking}), retain the top (1 - x) / 2 and discard the bottom (1 - x) /
* 2, and perform inference to evaluate the candidates in between
*/
protected static final float PARSE_FRACTION =
GlobalConfigProperties.singleton()
.getFloatProperty(PROPERTY_PARSE_FRACTION, DEFAULT_PARSE_FRACTION);
@SuppressWarnings("hiding")
public void init(
final CompleteClosureModel ccModel,
final List<String> trainingCorpus,
final List<String> developmentSet,
final float minRuleProbability) {
// Store CC model, training corpus, and dev-set (they'll be consistent throughout all training
// cycles)
this.ccModel = ccModel;
this.minRuleProbability = minRuleProbability;
this.ccModel = ccModel;
this.trainingCorpus = trainingCorpus;
// Un-binarize the dev-set
this.developmentSet = new ArrayList<String>();
for (final String binarizedTree : developmentSet) {
this.developmentSet.add(
BinaryTree.read(binarizedTree, String.class)
.unfactor(GrammarFormatType.Berkeley)
.toString());
}
}
/**
* Evaluates speed and accuracy with the baseline (fully-split) grammar
*
* @param grammar
* @param lexicon
* @param cycle
*/
@Override
public void initMergeCycle(final Grammar grammar, final Lexicon lexicon, final int cycle) {
this.splitGrammar = grammar;
this.splitLexicon = lexicon;
final String[] split =
GlobalConfigProperties.singleton()
.getProperty(PROPERTY_BEAM_WIDTHS, DEFAULT_BEAM_WIDTHS)
.split(",");
this.beamWidth = Integer.parseInt(split[cycle - 1]);
// Convert the grammar to BUBS sparse-matrix format and train a Boundary POS FOM
BaseLogger.singleton()
.info("Constrained parsing the training-set and training a prioritization model");
final LeftCscSparseMatrixGrammar sparseMatrixGrammar =
convertGrammarToSparseMatrix(splitGrammar, splitLexicon);
final BoundaryPosModel posFom = trainPosFom(sparseMatrixGrammar);
// Record accuracy with the full split grammar (so we can later compare with MergeCandidates)
BaseLogger.singleton().info("Parsing the dev-set with the fully-split grammar");
final float[] parseResult = parseDevSet(sparseMatrixGrammar, posFom, beamWidth);
splitF1 = parseResult[0];
splitSpeed = parseResult[1];
BaseLogger.singleton()
.info(String.format("F1 = %.3f Speed = %.3f w/s", splitF1 * 100, splitSpeed));
}
/**
* Parses the development set with the specified grammar and FOM. Returns the accuracy (F1) of the
* resulting parses. Uses the specified <code>beamWidth</code> for all cells spanning >= 2 words.
* For lexical cells, uses <code>beamWidth</code> x 3 (since the FOM does not prioritize entries
* in lexical cells), and allocates <code>beamWidth</code> entries in lexical cells for unary
* productions.
*
* @param sparseMatrixGrammar
* @param posFom
* @param cycle
* @return Accuracy (F1) and speed (w/s)
*/
protected float[] parseDevSet(
final LeftCscSparseMatrixGrammar sparseMatrixGrammar,
final BoundaryPosModel posFom,
final int cycle) {
// Initialize the parser
final ParserDriver opts = new ParserDriver();
opts.researchParserType = ResearchParserType.CartesianProductHashMl;
opts.cellSelectorModel = ccModel;
opts.fomModel = posFom;
// Set beam-width configuration properties
GlobalConfigProperties.singleton()
.setProperty(Parser.PROPERTY_MAX_BEAM_WIDTH, Integer.toString(beamWidth));
GlobalConfigProperties.singleton()
.setProperty(Parser.PROPERTY_LEXICAL_ROW_BEAM_WIDTH, Integer.toString(beamWidth * 3));
GlobalConfigProperties.singleton()
.setProperty(Parser.PROPERTY_LEXICAL_ROW_UNARIES, Integer.toString(beamWidth));
// Parse the dev-set
final CartesianProductHashSpmlParser parser =
new CartesianProductHashSpmlParser(opts, sparseMatrixGrammar);
final long t0 = System.currentTimeMillis();
int words = 0;
final BracketEvaluator evaluator = new BracketEvaluator();
for (final String inputTree : developmentSet) {
parser.parseSentence(inputTree).evaluate(evaluator);
final NaryTree<String> naryTree = NaryTree.read(inputTree, String.class);
words += naryTree.leaves();
}
final long t1 = System.currentTimeMillis();
return new float[] {(float) evaluator.accumulatedResult().f1(), words * 1000f / (t1 - t0)};
}
/**
* Converts a {@link Grammar} and {@link Lexicon} to BUBS sparse-matrix format, specifically
* {@link LeftCscSparseMatrixGrammar}. Prunes rules below the minimum rule probability threshold
* specified when the {@link DiscriminativeMergeObjectiveFunction} was initialized with {@link
* #init(CompleteClosureModel, List, List, float)}.
*
* @param grammar
* @param lexicon
* @return {@link LeftCscSparseMatrixGrammar}
*/
protected LeftCscSparseMatrixGrammar convertGrammarToSparseMatrix(
final Grammar grammar, final Lexicon lexicon) {
try {
final Writer w = new StringWriter(150 * 1024 * 1024);
// Note We could use a PipedOutputStream / PipedInputStream combination (with 2 threads) to
// write and read
// at the same time, and avoid using enough memory to serialize the entire grammar. But memory
// isn't a huge
// constraint during training, and the threading would add complexity, so we'll skip that for
// now.
w.write(grammar.toString(lexicon.totalRules(minRuleProbability), minRuleProbability, 0, 0));
w.write("===== LEXICON =====\n");
w.write(lexicon.toString(minRuleProbability));
return new LeftCscSparseMatrixGrammar(
new StringReader(w.toString()), new DecisionTreeTokenClassifier());
} catch (final IOException e) {
// StringWriter and StringReader should never IOException
throw new AssertionError(e);
}
}
/**
* Trains a boundary POS prioritization model (AKA a figure-of-merit, or FOM). Parses the training
* corpus, constrained by the gold trees, and learns prioritization probabilities from the
* resulting parses.
*
* @param sparseMatrixGrammar
* @return a boundary POS figure of merit model.
*/
protected BoundaryPosModel trainPosFom(final LeftCscSparseMatrixGrammar sparseMatrixGrammar) {
try {
// Constrained parse the training corpus
final ParserDriver opts = new ParserDriver();
opts.cellSelectorModel = ConstrainedCellSelector.MODEL;
opts.researchParserType = ResearchParserType.ConstrainedCartesianProductHashMl;
final ConstrainedCphSpmlParser constrainedParser =
new ConstrainedCphSpmlParser(opts, sparseMatrixGrammar);
final StringWriter binaryConstrainedParses = new StringWriter(30 * 1024 * 1024);
for (final String inputTree : trainingCorpus) {
final ParseTask parseTask = constrainedParser.parseSentence(inputTree);
binaryConstrainedParses.write(parseTask.binaryParse.toString());
binaryConstrainedParses.write('\n');
}
final StringWriter serializedFomModel = new StringWriter(30 * 1024 * 1024);
BoundaryPosModel.train(
sparseMatrixGrammar,
new BufferedReader(new StringReader(binaryConstrainedParses.toString())),
new BufferedWriter(serializedFomModel),
.5f,
false,
2);
final BufferedReader fomModelReader =
new BufferedReader(new StringReader(serializedFomModel.toString()));
return new BoundaryPosModel(FOMType.BoundaryPOS, sparseMatrixGrammar, fomModelReader);
} catch (final IOException e) {
// StringWriter and StringReader should never IOException
throw new AssertionError(e);
}
}
}