You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@joshua.apache.org by mj...@apache.org on 2016/04/22 06:17:49 UTC
[03/13] incubator-joshua git commit: Added full-file training,
start of feature function
Added full-file training, start of feature function
Project: http://git-wip-us.apache.org/repos/asf/incubator-joshua/repo
Commit: http://git-wip-us.apache.org/repos/asf/incubator-joshua/commit/47f1af58
Tree: http://git-wip-us.apache.org/repos/asf/incubator-joshua/tree/47f1af58
Diff: http://git-wip-us.apache.org/repos/asf/incubator-joshua/diff/47f1af58
Branch: refs/heads/morph
Commit: 47f1af588a8878963112b11ef29e4180e9d82869
Parents: dca7a5d
Author: Matt Post <po...@cs.jhu.edu>
Authored: Thu Apr 21 09:15:09 2016 -0400
Committer: Matt Post <po...@cs.jhu.edu>
Committed: Thu Apr 21 09:15:09 2016 -0400
----------------------------------------------------------------------
.../decoder/ff/morph/InflectionPredictor.java | 193 ++++++++++++++++---
src/joshua/decoder/segment_file/Token.java | 12 +-
2 files changed, 178 insertions(+), 27 deletions(-)
----------------------------------------------------------------------
http://git-wip-us.apache.org/repos/asf/incubator-joshua/blob/47f1af58/src/joshua/decoder/ff/morph/InflectionPredictor.java
----------------------------------------------------------------------
diff --git a/src/joshua/decoder/ff/morph/InflectionPredictor.java b/src/joshua/decoder/ff/morph/InflectionPredictor.java
index 5282497..f4a4310 100644
--- a/src/joshua/decoder/ff/morph/InflectionPredictor.java
+++ b/src/joshua/decoder/ff/morph/InflectionPredictor.java
@@ -1,25 +1,40 @@
package joshua.decoder.ff.morph;
-/*
+/***
+ * This feature function scores a rule application by predicting, for each target word aligned with
+ * a source word, how likely the lexical translation is in context.
+ *
+ * The feature function can be provided with a trained model or a raw training file which it will
+ * then train prior to decoding.
+ *
* Format of training file:
*
* source_word target_word feature:value feature:value feature:value ...
*
* Invocation:
*
- * java -cp /Users/post/code/joshua/lib/mallet-2.0.7.jar:/Users/post/code/joshua/lib/trove4j-2.0.2.jar:$JOSHUA/class joshua.decoder.ff.morph.InflectionPredictor /path/to/training/data
+ * java -cp /Users/post/code/joshua/lib/mallet-2.0.7.jar:/Users/post/code/joshua/lib/trove4j-2.0.2.jar:$JOSHUA/class joshua.decoder.ff.morph.LexicalSharpener /path/to/training/data
*/
import java.io.File;
+import java.io.FileInputStream;
import java.io.FileNotFoundException;
+import java.io.FileOutputStream;
import java.io.FileReader;
+import java.io.IOException;
+import java.io.ObjectInputStream;
+import java.io.ObjectOutputStream;
import java.util.ArrayList;
import java.util.List;
+import java.util.Map;
+import java.util.Scanner;
import cc.mallet.classify.*;
import cc.mallet.pipe.*;
import cc.mallet.pipe.iterator.CsvIterator;
+import cc.mallet.types.Instance;
import cc.mallet.types.InstanceList;
+import joshua.corpus.Vocabulary;
import joshua.decoder.JoshuaConfiguration;
import joshua.decoder.chart_parser.SourcePath;
import joshua.decoder.ff.FeatureVector;
@@ -28,13 +43,31 @@ import joshua.decoder.ff.state_maintenance.DPState;
import joshua.decoder.ff.tm.Rule;
import joshua.decoder.hypergraph.HGNode;
import joshua.decoder.segment_file.Sentence;
+import joshua.decoder.segment_file.Token;
public class InflectionPredictor extends StatelessFF {
private Classifier classifier = null;
+ private SerialPipes pipes = null;
public InflectionPredictor(final FeatureVector weights, String[] args, JoshuaConfiguration config) {
- super(weights, "InfectionPredictor", args, config);
+ super(weights, "LexicalSharpener", args, config);
+
+ ArrayList<Pipe> pipeList = new ArrayList<Pipe>();
+
+ // I don't know if this is needed
+ pipeList.add(new Target2Label());
+ // Convert SVM-light format to sparse feature vector
+ pipeList.add(new SvmLight2FeatureVectorAndLabel());
+ // Validation
+// pipeList.add(new PrintInputAndTarget());
+
+ // name: english word
+ // data: features (FeatureVector)
+ // target: foreign inflection
+ // source: null
+
+ pipes = new SerialPipes(pipeList);
if (parsedArgs.containsKey("model")) {
String modelFile = parsedArgs.get("model");
@@ -51,28 +84,30 @@ public class InflectionPredictor extends StatelessFF {
System.exit(1);
}
} else {
- // TODO: load the model
+ try {
+ loadClassifier(modelFile);
+ } catch (IOException e) {
+ // TODO Auto-generated catch block
+ e.printStackTrace();
+ } catch (ClassNotFoundException e) {
+ // TODO Auto-generated catch block
+ e.printStackTrace();
+ }
}
}
}
+ /**
+ * Trains a maxent classifier from the provided training data, returning a Mallet model.
+ *
+ * @param dataFile
+ * @return
+ * @throws FileNotFoundException
+ */
public Classifier train(String dataFile) throws FileNotFoundException {
- ArrayList<Pipe> pipeList = new ArrayList<Pipe>();
-
- // I don't know if this is needed
- pipeList.add(new Target2Label());
- // Convert SVM-light format to sparse feature vector
- pipeList.add(new SvmLight2FeatureVectorAndLabel());
- // Validation
-// pipeList.add(new PrintInputAndTarget());
-
- // name: english word
- // data: features (FeatureVector)
- // target: foreign inflection
- // source: null
// Remove the first field (Mallet's "name" field), leave the rest for SVM-light conversion
- InstanceList instances = new InstanceList(new SerialPipes(pipeList));
+ InstanceList instances = new InstanceList(pipes);
instances.addThruPipe(new CsvIterator(new FileReader(dataFile),
"(\\w+)\\s+(.*)",
2, -1, 1));
@@ -82,22 +117,130 @@ public class InflectionPredictor extends StatelessFF {
return classifier;
}
-
+
+ public void loadClassifier(String modelFile) throws ClassNotFoundException, IOException {
+ ObjectInputStream ois = new ObjectInputStream(new FileInputStream(modelFile));
+ classifier = (Classifier) ois.readObject();
+ }
+
+ public void saveClassifier(String modelFile) throws FileNotFoundException, IOException {
+ ObjectOutputStream oos = new ObjectOutputStream(new FileOutputStream(modelFile));
+ oos.writeObject(classifier);
+ oos.close();
+ }
+
+ public Classification predict(String outcome, String features) {
+ Instance instance = new Instance(features, null, null, null);
+ System.err.println("PREDICT outcome = " + (String) instance.getTarget());
+ System.err.println("PREDICT features = " + (String) instance.getData());
+ Classification result = (Classification) classifier.classify(pipes.instanceFrom(instance));
+
+ return result;
+ }
+
+ /**
+ * Compute features. This works by walking over the target side phrase pieces, looking for every
+ * word with a single source-aligned word. We then throw the annotations from that source word
+ * into our prediction model to learn how much it likes the chosen word. Presumably the source-
+ * language annotations have contextual features, so this effectively chooses the words in context.
+ */
@Override
public DPState compute(Rule rule, List<HGNode> tailNodes, int i, int j, SourcePath sourcePath,
Sentence sentence, Accumulator acc) {
+
+ Map<Integer, List<Integer>> points = rule.getAlignmentMap();
+ for (int t: points.keySet()) {
+ List<Integer> source_indices = points.get(t);
+ if (source_indices.size() != 1)
+ continue;
+
+ String targetWord = Vocabulary.word(rule.getEnglish()[t]);
+ int s = i + source_indices.get(0);
+ Token sourceToken = sentence.getTokens().get(s);
+ String featureString = sourceToken.getAnnotationString().replace('|', ' ');
+
+ Classification result = predict(targetWord, featureString);
+ if (result.bestLabelIsCorrect()) {
+ acc.add(String.format("%s_match", name), 1);
+ }
+ }
return null;
}
- public static void main(String[] args) throws FileNotFoundException {
- InflectionPredictor ip = new InflectionPredictor(null, args, null);
+ /**
+ * Returns an array parallel to the source words array indicating, for each index, the absolute
+ * position of that word into the source sentence. For example, for the rule with source side
+ *
+ * [ 17, 142, -14, 9 ]
+ *
+ * and source sentence
+ *
+ * [ 17, 18, 142, 1, 1, 9, 8 ]
+ *
+ * it will return
+ *
+ * [ 0, 2, -14, 5 ]
+ *
+ * which indicates that the first, second, and fourth words of the rule are anchored to the
+ * first, third, and sixth words of the input sentence.
+ *
+ * @param rule
+ * @param tailNodes
+ * @param start
+ * @return a list of alignment points anchored to the source sentence
+ */
+ public int[] anchorRuleSourceToSentence(Rule rule, List<HGNode> tailNodes, int start) {
+ int[] source = rule.getFrench();
+
+ // Map the source words in the rule to absolute positions in the sentence
+ int[] anchoredSource = source.clone();
- String dataFile = "/Users/post/Desktop/amazon16/model";
- if (args.length > 0)
- dataFile = args[0];
+ int sourceIndex = start;
+ int tailNodeIndex = 0;
+ for (int i = 0; i < source.length; i++) {
+ if (source[i] < 0) { // nonterminal
+ anchoredSource[i] = source[i];
+ sourceIndex = tailNodes.get(tailNodeIndex).j;
+ tailNodeIndex++;
+ } else { // terminal
+ anchoredSource[i] = sourceIndex;
+ sourceIndex++;
+ }
+ }
- ip.train(dataFile);
+ return anchoredSource;
}
+ public static void main(String[] args) throws IOException, ClassNotFoundException {
+ InflectionPredictor ts = new InflectionPredictor(null, args, null);
+
+ String modelFile = "model";
+
+ if (args.length > 0) {
+ String dataFile = args[0];
+
+ System.err.println("Training model from file " + dataFile);
+ ts.train(dataFile);
+
+ if (args.length > 1)
+ modelFile = args[1];
+
+ System.err.println("Writing model to file " + modelFile);
+ ts.saveClassifier(modelFile);
+ } else {
+ System.err.println("Loading model from file " + modelFile);
+ ts.loadClassifier(modelFile);
+ }
+
+ Scanner stdin = new Scanner(System.in);
+ while(stdin.hasNextLine()) {
+ String line = stdin.nextLine();
+ String[] tokens = line.split(" ", 2);
+ String outcome = tokens[0];
+ String features = tokens[1];
+ Classification result = ts.predict(outcome, features);
+ System.out.println(String.format("%s %f", result.getLabelVector().getBestLabel(), result.getLabelVector().getBestValue()));
+ }
+ }
}
http://git-wip-us.apache.org/repos/asf/incubator-joshua/blob/47f1af58/src/joshua/decoder/segment_file/Token.java
----------------------------------------------------------------------
diff --git a/src/joshua/decoder/segment_file/Token.java b/src/joshua/decoder/segment_file/Token.java
index 12e2b68..7969294 100644
--- a/src/joshua/decoder/segment_file/Token.java
+++ b/src/joshua/decoder/segment_file/Token.java
@@ -36,6 +36,7 @@ public class Token {
private int tokenID;
private HashMap<String,String> annotations = null;
+ private String annotationString;
/**
* Constructor : Creates a Token object from a raw word
@@ -69,9 +70,9 @@ public class Token {
if (tag.find()) {
// Annotation match found
token = tag.group(1);
- String tagStr = tag.group(2);
+ annotationString = tag.group(2);
- for (String annotation: tagStr.split(";")) {
+ for (String annotation: annotationString.split(";")) {
int where = annotation.indexOf("=");
if (where != -1) {
annotations.put(annotation.substring(0, where), annotation.substring(where + 1));
@@ -121,4 +122,11 @@ public class Token {
return null;
}
+
+ /**
+ * Returns the raw annotation string
+ */
+ public String getAnnotationString() {
+ return annotationString;
+ }
}
\ No newline at end of file