-
Notifications
You must be signed in to change notification settings - Fork 2.7k
/
ScorePhrasesLearnFeatWt.java
1414 lines (1156 loc) · 60.4 KB
/
ScorePhrasesLearnFeatWt.java
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
package edu.stanford.nlp.patterns;
import java.io.BufferedWriter;
import java.io.File;
import java.io.FileWriter;
import java.io.IOException;
import java.util.*;
import java.util.Map.Entry;
import java.util.concurrent.*;
import java.util.function.Predicate;
import java.util.stream.Collectors;
import edu.stanford.nlp.classify.*;
import edu.stanford.nlp.io.IOUtils;
import edu.stanford.nlp.ling.BasicDatum;
import edu.stanford.nlp.ling.CoreLabel;
import edu.stanford.nlp.ling.IndexedWord;
import edu.stanford.nlp.ling.RVFDatum;
import edu.stanford.nlp.patterns.ConstantsAndVariables.ScorePhraseMeasures;
import edu.stanford.nlp.patterns.dep.DataInstanceDep;
import edu.stanford.nlp.patterns.dep.ExtractPhraseFromPattern;
import edu.stanford.nlp.patterns.dep.ExtractedPhrase;
import edu.stanford.nlp.semgraph.SemanticGraph;
import edu.stanford.nlp.stats.*;
import edu.stanford.nlp.util.*;
import edu.stanford.nlp.util.ArgumentParser.Option;
import edu.stanford.nlp.util.concurrent.AtomicDouble;
import edu.stanford.nlp.util.concurrent.ConcurrentHashCounter;
import edu.stanford.nlp.util.logging.Redwood;
/**
* Learn a logistic regression classifier to combine weights to score a phrase.
*
* @author Sonal Gupta (sonalg@stanford.edu)
*
*/
public class ScorePhrasesLearnFeatWt<E extends Pattern> extends PhraseScorer<E> {
@Option(name = "scoreClassifierType")
private ClassifierType scoreClassifierType = ClassifierType.LR;
private static Map<String, double[]> wordVectors = null;
public ScorePhrasesLearnFeatWt(ConstantsAndVariables constvar) {
super(constvar);
if(constvar.useWordVectorsToComputeSim && (constvar.subsampleUnkAsNegUsingSim|| constvar.expandPositivesWhenSampling || constvar.expandNegativesWhenSampling || constVars.usePhraseEvalWordVector) && wordVectors == null) {
if(Data.rawFreq == null){
Data.rawFreq = new ClassicCounter<>();
Data.computeRawFreqIfNull(PatternFactory.numWordsCompoundMax, constvar.batchProcessSents);
}
Redwood.log(Redwood.DBG, "Reading word vectors");
wordVectors = new HashMap<>();
for (String line : IOUtils.readLines(constVars.wordVectorFile)) {
String[] tok = line.split("\\s+");
String word = tok[0];
CandidatePhrase p = CandidatePhrase.createOrGet(word);
//save the vector if it occurs in the rawFreq, seed set, stop words, english words
if (Data.rawFreq.containsKey(p) || constvar.getStopWords().contains(p) || constvar.getEnglishWords().contains(word) || constvar.hasSeedWordOrOtherSem(p)) {
double[] d = new double[tok.length - 1];
for (int i = 1; i < tok.length; i++) {
d[i - 1] = Double.valueOf(tok[i]);
}
wordVectors.put(word, d);
} else
CandidatePhrase.deletePhrase(p);
}
Redwood.log(Redwood.DBG, "Read " + wordVectors.size() + " word vectors");
}
OOVExternalFeatWt = 0;
OOVdictOdds = 0;
OOVDomainNgramScore = 0;
OOVGoogleNgramScore = 0;
}
public enum ClassifierType {
DT, LR, RF, SVM, SHIFTLR, LINEAR
}
public TwoDimensionalCounter<CandidatePhrase, ScorePhraseMeasures> phraseScoresRaw = new TwoDimensionalCounter<>();
public edu.stanford.nlp.classify.Classifier learnClassifier(String label, boolean forLearningPatterns,
TwoDimensionalCounter<CandidatePhrase, E> wordsPatExtracted, Counter<E> allSelectedPatterns) throws IOException, ClassNotFoundException {
phraseScoresRaw.clear();
learnedScores.clear();
if(Data.domainNGramsFile != null)
Data.loadDomainNGrams();
boolean computeRawFreq = false;
if (Data.rawFreq == null) {
Data.rawFreq = new ClassicCounter<>();
computeRawFreq = true;
}
GeneralDataset<String, ScorePhraseMeasures> dataset = choosedatums(forLearningPatterns, label, wordsPatExtracted, allSelectedPatterns, computeRawFreq);
edu.stanford.nlp.classify.Classifier classifier;
if (scoreClassifierType.equals(ClassifierType.LR)) {
LogisticClassifierFactory<String, ScorePhraseMeasures> logfactory = new LogisticClassifierFactory<>();
LogPrior lprior = new LogPrior();
lprior.setSigma(constVars.LRSigma);
classifier = logfactory.trainClassifier(dataset, lprior, false);
LogisticClassifier logcl = ((LogisticClassifier) classifier);
String l = (String) logcl.getLabelForInternalPositiveClass();
Counter<String> weights = logcl.weightsAsCounter();
if (l.equals(Boolean.FALSE.toString())) {
Counters.multiplyInPlace(weights, -1);
}
List<Pair<String, Double>> wtd = Counters.toDescendingMagnitudeSortedListWithCounts(weights);
Redwood.log(ConstantsAndVariables.minimaldebug, "The weights are " + StringUtils.join(wtd.subList(0, Math.min(wtd.size(), 600)), "\n"));
} else if(scoreClassifierType.equals(ClassifierType.SVM)){
SVMLightClassifierFactory<String, ScorePhraseMeasures> svmcf = new SVMLightClassifierFactory<>(true);
classifier = svmcf.trainClassifier(dataset);
Set<String> labels = Generics.newHashSet(Arrays.asList("true"));
List<Triple<ScorePhraseMeasures, String, Double>> topfeatures = ((SVMLightClassifier<String, ScorePhraseMeasures>) classifier).getTopFeatures(labels, 0, true, 600, true);
Redwood.log(ConstantsAndVariables.minimaldebug, "The weights are " + StringUtils.join(topfeatures, "\n"));
}else if(scoreClassifierType.equals(ClassifierType.SHIFTLR)){
//change the dataset to basic dataset because currently ShiftParamsLR doesn't support RVFDatum
GeneralDataset<String, ScorePhraseMeasures> newdataset = new Dataset<>();
Iterator<RVFDatum<String, ScorePhraseMeasures>> iter = dataset.iterator();
while(iter.hasNext()){
RVFDatum<String, ScorePhraseMeasures> inst = iter.next();
newdataset.add(new BasicDatum<>(inst.asFeatures(), inst.label()));
}
ShiftParamsLogisticClassifierFactory<String, ScorePhraseMeasures> factory = new ShiftParamsLogisticClassifierFactory<>();
classifier = factory.trainClassifier(newdataset);
//print weights
MultinomialLogisticClassifier<String, ScorePhraseMeasures> logcl = ((MultinomialLogisticClassifier) classifier);
Counter<ScorePhraseMeasures> weights = logcl.weightsAsGenericCounter().get("true");
List<Pair<ScorePhraseMeasures, Double>> wtd = Counters.toDescendingMagnitudeSortedListWithCounts(weights);
Redwood.log(ConstantsAndVariables.minimaldebug, "The weights are " + StringUtils.join(wtd.subList(0, Math.min(wtd.size(), 600)), "\n"));
} else if(scoreClassifierType.equals(ClassifierType.LINEAR)){
LinearClassifierFactory<String, ScorePhraseMeasures> lcf = new LinearClassifierFactory<>();
classifier = lcf.trainClassifier(dataset);
Set<String> labels = Generics.newHashSet(Arrays.asList("true"));
List<Triple<ScorePhraseMeasures, String, Double>> topfeatures = ((LinearClassifier<String, ScorePhraseMeasures>) classifier).getTopFeatures(labels, 0, true, 600, true);
Redwood.log(ConstantsAndVariables.minimaldebug, "The weights are " + StringUtils.join(topfeatures, "\n"));
}else
throw new RuntimeException("cannot identify classifier " + scoreClassifierType);
// else if (scoreClassifierType.equals(ClassifierType.RF)) {
// ClassifierFactory wekaFactory = new WekaDatumClassifierFactory<String, ScorePhraseMeasures>("weka.classifiers.trees.RandomForest", constVars.wekaOptions);
// classifier = wekaFactory.trainClassifier(dataset);
// Classifier cls = ((WekaDatumClassifier) classifier).getClassifier();
// RandomForest rf = (RandomForest) cls;
// }
BufferedWriter w = new BufferedWriter(new FileWriter("tempscorestrainer.txt"));
System.out.println("size of learned scores is " + phraseScoresRaw.size());
for (CandidatePhrase s : phraseScoresRaw.firstKeySet()) {
w.write(s + "\t" + phraseScoresRaw.getCounter(s) + "\n");
}
w.close();
return classifier;
}
@Override
public void printReasonForChoosing(Counter<CandidatePhrase> phrases){
Redwood.log(Redwood.DBG, "Features of selected phrases");
for(Entry<CandidatePhrase, Double> pEn: phrases.entrySet())
Redwood.log(Redwood.DBG, pEn.getKey().getPhrase() + "\t" + pEn.getValue() + "\t" + phraseScoresRaw.getCounter(pEn.getKey()));
}
@Override
public Counter<CandidatePhrase> scorePhrases(String label, TwoDimensionalCounter<CandidatePhrase, E> terms,
TwoDimensionalCounter<CandidatePhrase, E> wordsPatExtracted, Counter<E> allSelectedPatterns,
Set<CandidatePhrase> alreadyIdentifiedWords, boolean forLearningPatterns) throws IOException, ClassNotFoundException {
getAllLabeledWordsCluster();
Counter<CandidatePhrase> scores = new ClassicCounter<>();
edu.stanford.nlp.classify.Classifier classifier = learnClassifier(label, forLearningPatterns, wordsPatExtracted, allSelectedPatterns);
for (Entry<CandidatePhrase, ClassicCounter<E>> en : terms.entrySet()) {
Double score = this.scoreUsingClassifer(classifier, en.getKey(), label, forLearningPatterns, en.getValue(), allSelectedPatterns);
if(!score.isNaN() && !score.isInfinite()){
scores.setCount(en.getKey(), score);
}else
Redwood.log(Redwood.DBG, "Ignoring " + en.getKey() + " because score is " + score);
}
return scores;
}
@Override
public Counter<CandidatePhrase> scorePhrases(String label, Set<CandidatePhrase> terms, boolean forLearningPatterns) throws IOException, ClassNotFoundException {
getAllLabeledWordsCluster();
Counter<CandidatePhrase> scores = new ClassicCounter<>();
edu.stanford.nlp.classify.Classifier classifier = learnClassifier(label, forLearningPatterns, null, null);
for (CandidatePhrase en : terms) {
double score = this.scoreUsingClassifer(classifier, en, label, forLearningPatterns,null, null);
scores.setCount(en, score);
}
return scores;
}
public static boolean getRandomBoolean(Random random, double p) {
return random.nextFloat() < p;
}
static double logistic(double d) {
return 1 / (1 + Math.exp(-1 * d));
}
ConcurrentHashMap<CandidatePhrase, Counter<Integer>> wordClassClustersForPhrase = new ConcurrentHashMap<>();
Counter<Integer> wordClass(String phrase, String phraseLemma){
Counter<Integer> cl = new ClassicCounter<>();
String[] phl = null;
if(phraseLemma!=null)
phl = phraseLemma.split("\\s+");
int i =0;
for(String w: phrase.split("\\s+")) {
Integer cluster = constVars.getWordClassClusters().get(w);
if (cluster == null && phl!=null)
cluster = constVars.getWordClassClusters().get(phl[i]);
//try lowercase
if(cluster == null){
cluster = constVars.getWordClassClusters().get(w.toLowerCase());
if (cluster == null && phl!=null)
cluster = constVars.getWordClassClusters().get(phl[i].toLowerCase());
}
if(cluster != null)
cl.incrementCount(cluster);
i++;
}
return cl;
}
void getAllLabeledWordsCluster(){
for(String label: constVars.getLabels()){
for(Map.Entry<CandidatePhrase, Double> p : constVars.getLearnedWords(label).entrySet()){
wordClassClustersForPhrase.put(p.getKey(), wordClass(p.getKey().getPhrase(), p.getKey().getPhraseLemma()));
}
for(CandidatePhrase p : constVars.getSeedLabelDictionary().get(label)){
wordClassClustersForPhrase.put(p, wordClass(p.getPhrase(), p.getPhraseLemma()));
}
}
}
private Counter<CandidatePhrase> computeSimWithWordVectors(Collection<CandidatePhrase> candidatePhrases, Collection<CandidatePhrase> otherPhrases, boolean ignoreWordRegex, String label){
Counter<CandidatePhrase> sims = new ClassicCounter<>(candidatePhrases.size());
for(CandidatePhrase p : candidatePhrases) {
Map<String, double[]> simsAvgMaxAllLabels = similaritiesWithLabeledPhrases.get(p.getPhrase());
if(simsAvgMaxAllLabels == null)
simsAvgMaxAllLabels = new HashMap<>();
double[] simsAvgMax = simsAvgMaxAllLabels.get(label);
if (simsAvgMax == null) {
simsAvgMax = new double[Similarities.values().length];
// Arrays.fill(simsAvgMax, 0); // not needed; Java arrays zero initialized
}
if(wordVectors.containsKey(p.getPhrase()) && (! ignoreWordRegex || !PatternFactory.ignoreWordRegex.matcher(p.getPhrase()).matches())){
double[] d1 = wordVectors.get(p.getPhrase());
BinaryHeapPriorityQueue<CandidatePhrase> topSimPhs = new BinaryHeapPriorityQueue<>(constVars.expandPhrasesNumTopSimilar);
double allsum = 0;
double max = Double.MIN_VALUE;
boolean donotuse = false;
for (CandidatePhrase other : otherPhrases) {
if (p.equals(other)) {
donotuse = true;
break;
}
if (!wordVectors.containsKey(other.getPhrase()))
continue;
double sim;
PhrasePair pair = new PhrasePair(p.getPhrase(), other.getPhrase());
if (cacheSimilarities.containsKey(pair))
sim = cacheSimilarities.getCount(pair);
else {
double[] d2 = wordVectors.get(other.getPhrase());
double sum = 0;
double d1sq = 0;
double d2sq = 0;
for (int i = 0; i < d1.length; i++) {
sum += d1[i] * d2[i];
d1sq += d1[i] * d1[i];
d2sq += d2[i] * d2[i];
}
sim = sum / (Math.sqrt(d1sq) * Math.sqrt(d2sq));
cacheSimilarities.setCount(pair, sim);
}
topSimPhs.add(other, sim);
if(topSimPhs.size() > constVars.expandPhrasesNumTopSimilar)
topSimPhs.removeLastEntry();
//avgSim /= otherPhrases.size();
allsum += sim;
if(sim > max)
max = sim;
}
double finalSimScore = 0;
int numEl = 0;
while(topSimPhs.hasNext()) {
finalSimScore += topSimPhs.getPriority();
topSimPhs.next();
numEl++;
}
finalSimScore /= numEl;
double prevNumItems = simsAvgMax[Similarities.NUMITEMS.ordinal()];
double prevAvg = simsAvgMax[Similarities.AVGSIM.ordinal()];
double prevMax = simsAvgMax[Similarities.MAXSIM.ordinal()];
double newNumItems = prevNumItems + otherPhrases.size();
double newAvg = (prevAvg*prevNumItems + allsum) /(newNumItems);
double newMax = prevMax > max ? prevMax: max;
simsAvgMax[Similarities.NUMITEMS.ordinal()] = newNumItems;
simsAvgMax[Similarities.AVGSIM.ordinal()] = newAvg;
simsAvgMax[Similarities.MAXSIM.ordinal()] = newMax;
if(!donotuse){
sims.setCount(p, finalSimScore);
}
}else{
sims.setCount(p, Double.MIN_VALUE);
}
simsAvgMaxAllLabels.put(label, simsAvgMax);
similaritiesWithLabeledPhrases.put(p.getPhrase(), simsAvgMaxAllLabels);
}
return sims;
}
private Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>> computeSimWithWordVectors(List<CandidatePhrase> candidatePhrases, Collection<CandidatePhrase> positivePhrases,
Map<String, Collection<CandidatePhrase>> allPossibleNegativePhrases, String label) {
assert wordVectors != null : "Why are word vectors null?";
Counter<CandidatePhrase> posSims = computeSimWithWordVectors(candidatePhrases, positivePhrases, true, label);
Counter<CandidatePhrase> negSims = new ClassicCounter<>();
for(Map.Entry<String, Collection<CandidatePhrase>> en: allPossibleNegativePhrases.entrySet())
negSims.addAll(computeSimWithWordVectors(candidatePhrases, en.getValue(), true, en.getKey()));
Predicate<CandidatePhrase> retainPhrasesNotCloseToNegative = candidatePhrase -> {
if(negSims.getCount(candidatePhrase) > posSims.getCount(candidatePhrase))
return false;
else
return true;
};
Counters.retainKeys(posSims, retainPhrasesNotCloseToNegative);
return new Pair(posSims, negSims);
}
Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>> computeSimWithWordCluster(Collection<CandidatePhrase> candidatePhrases, Collection<CandidatePhrase> positivePhrases, AtomicDouble allMaxSim){
Counter<CandidatePhrase> sims = new ClassicCounter<>(candidatePhrases.size());
for(CandidatePhrase p : candidatePhrases) {
Counter<Integer> feat = wordClassClustersForPhrase.get(p);
if(feat == null){
feat = wordClass(p.getPhrase(), p.getPhraseLemma());
wordClassClustersForPhrase.put(p, feat);
}
double avgSim = 0;// Double.MIN_VALUE;
if(feat.size() > 0) {
for (CandidatePhrase pos : positivePhrases) {
if(p.equals(pos))
continue;
Counter<Integer> posfeat = wordClassClustersForPhrase.get(pos);
if(posfeat == null){
posfeat = wordClass(pos.getPhrase(), pos.getPhraseLemma());
wordClassClustersForPhrase.put(pos, feat);
}
if(posfeat.size() > 0){
Double j = Counters.jaccardCoefficient(posfeat, feat);
//System.out.println("clusters for positive phrase " + pos + " is " +wordClassClustersForPhrase.get(pos) + " and the features for unknown are " + feat + " for phrase " + p);
if(!j.isInfinite() && !j.isNaN()){
avgSim += j;
}
//if (j > maxSim)
// maxSim = j;
}
}
avgSim /= positivePhrases.size();
}
sims.setCount(p, avgSim);
if(allMaxSim.get() < avgSim)
allMaxSim.set(avgSim);
}
//TODO: compute similarity with neg phrases
return new Pair(sims, null);
}
class ComputeSim implements Callable<Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>>>{
List<CandidatePhrase> candidatePhrases;
String label;
AtomicDouble allMaxSim;
Collection<CandidatePhrase> positivePhrases;
Map<String, Collection<CandidatePhrase>> knownNegativePhrases;
public ComputeSim(String label, List<CandidatePhrase> candidatePhrases, AtomicDouble allMaxSim, Collection<CandidatePhrase> positivePhrases, Map<String, Collection<CandidatePhrase>> knownNegativePhrases){
this.label = label;
this.candidatePhrases = candidatePhrases;
this.allMaxSim = allMaxSim;
this.positivePhrases = positivePhrases;
this.knownNegativePhrases = knownNegativePhrases;
}
@Override
public Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>> call() throws Exception {
if(constVars.useWordVectorsToComputeSim){
Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>> phs = computeSimWithWordVectors(candidatePhrases, positivePhrases, knownNegativePhrases, label);
Redwood.log(Redwood.DBG, "Computed similarities with positive and negative phrases");
return phs;
}
else
//TODO: knownnegaitvephrases
return computeSimWithWordCluster(candidatePhrases, positivePhrases, allMaxSim);
}
}
//this chooses the ones that are not close to the positive phrases!
Set<CandidatePhrase> chooseUnknownAsNegatives(Set<CandidatePhrase> candidatePhrases, String label, Collection<CandidatePhrase> positivePhrases, Map<String,
Collection<CandidatePhrase>> knownNegativePhrases, BufferedWriter logFile) throws IOException {
List<List<CandidatePhrase>> threadedCandidates = GetPatternsFromDataMultiClass.getThreadBatches(CollectionUtils.toList(candidatePhrases), constVars.numThreads);
Counter<CandidatePhrase> sims = new ClassicCounter<>();
AtomicDouble allMaxSim = new AtomicDouble(Double.MIN_VALUE);
ExecutorService executor = Executors.newFixedThreadPool(constVars.numThreads);
List<Future<Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>>>> list = new ArrayList<>();
//multi-threaded choose positive, negative and unknown
for (List<CandidatePhrase> keys : threadedCandidates) {
Callable<Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>>> task = new ComputeSim(label, keys, allMaxSim, positivePhrases, knownNegativePhrases);
Future<Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>>> submit = executor.submit(task);
list.add(submit);
}
// Now retrieve the result
for (Future<Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>>> future : list) {
try {
sims.addAll(future.get().first());
} catch (Exception e) {
executor.shutdownNow();
throw new RuntimeException(e);
}
}
executor.shutdown();
if(allMaxSim.get() == Double.MIN_VALUE){
Redwood.log(Redwood.DBG, "No similarity recorded between the positives and the unknown!");
}
CandidatePhrase k = Counters.argmax(sims);
System.out.println("Maximum similarity was " + sims.getCount(k) + " for word " + k);
Counter<CandidatePhrase> removed = Counters.retainBelow(sims, constVars.positiveSimilarityThresholdLowPrecision);
System.out.println("removing phrases as negative phrases that were higher that positive similarity threshold of " + constVars.positiveSimilarityThresholdLowPrecision + removed);
if(logFile != null && wordVectors != null){
for(Entry<CandidatePhrase, Double> en: removed.entrySet())
if(wordVectors.containsKey(en.getKey().getPhrase()))
logFile.write(en.getKey()+"-PN " + ArrayUtils.toString(wordVectors.get(en.getKey().getPhrase()), " ")+"\n");
}
//Collection<CandidatePhrase> removed = Counters.retainBottom(sims, (int) (sims.size() * percentage));
//System.out.println("not choosing " + removed + " as the negative phrases. percentage is " + percentage + " and allMaxsim was " + allMaxSim);
return sims.keySet();
}
Set<CandidatePhrase> chooseUnknownPhrases(DataInstance sent, Random random, double perSelect, Class positiveClass, String label, int maxNum){
Set<CandidatePhrase> unknownSamples = new HashSet<>();
if(maxNum == 0)
return unknownSamples;
Predicate<CoreLabel> acceptWord = coreLabel -> {
if(coreLabel.get(positiveClass).equals(label) || constVars.functionWords.contains(coreLabel.word()))
return false;
else
return true;
};
Random r = new Random(0);
List<Integer> lengths = new ArrayList<>();
for(int i = 1;i <= PatternFactory.numWordsCompoundMapped.get(label); i++)
lengths.add(i);
int length = CollectionUtils.sample(lengths, r);
if(constVars.patternType.equals(PatternFactory.PatternType.DEP)){
ExtractPhraseFromPattern extract = new ExtractPhraseFromPattern(true, length);
SemanticGraph g = ((DataInstanceDep) sent).getGraph();
Collection<CoreLabel> sampledHeads = CollectionUtils.sampleWithoutReplacement(sent.getTokens(), Math.min(maxNum, (int) (perSelect * sent.getTokens().size())), random);
//TODO: change this for more efficient implementation
List<String> textTokens = sent.getTokens().stream().map(x -> x.word()).collect(Collectors.toList());
for(CoreLabel l: sampledHeads) {
if(!acceptWord.test(l))
continue;
IndexedWord w = g.getNodeByIndex(l.index());
List<String> outputPhrases = new ArrayList<>();
List<ExtractedPhrase> extractedPhrases = new ArrayList<>();
List<IntPair> outputIndices = new ArrayList<>();
extract.printSubGraph(g, w, new ArrayList<>(), textTokens, outputPhrases, outputIndices, new ArrayList<>(), new ArrayList<>(),
false, extractedPhrases, null, acceptWord);
for(ExtractedPhrase p :extractedPhrases){
unknownSamples.add(CandidatePhrase.createOrGet(p.getValue(), null, p.getFeatures()));
}
}
}else if(constVars.patternType.equals(PatternFactory.PatternType.SURFACE)){
CoreLabel[] tokens = sent.getTokens().toArray(new CoreLabel[0]);
for(int i =0; i < tokens.length; i++){
if(random.nextDouble() < perSelect){
int left = (int)((length -1) /2.0);
int right = length -1 -left;
String ph = "";
boolean haspositive = false;
for(int j = Math.max(0, i - left); j < tokens.length && j <= i+right; j++){
if(tokens[j].get(positiveClass).equals(label)){
haspositive = true;
break;
}
ph += " " + tokens[j].word();
}
ph = ph.trim();
if(!haspositive && !ph.trim().isEmpty() && !constVars.functionWords.contains(ph)){
unknownSamples.add(CandidatePhrase.createOrGet(ph));
}
}
}
} else
throw new RuntimeException("not yet implemented");
return unknownSamples;
}
private static<E,F> boolean hasElement(Map<E, Collection<F>> values, F value, E ignoreLabel){
for(Map.Entry<E, Collection<F>> en: values.entrySet()){
if(en.getKey().equals(ignoreLabel))
continue;
if(en.getValue().contains(value))
return true;
}
return false;
}
Counter<String> numLabeledTokens(){
Counter<String> counter = new ClassicCounter<>();
ConstantsAndVariables.DataSentsIterator data = new ConstantsAndVariables.DataSentsIterator(constVars.batchProcessSents);
while(data.hasNext()){
Map<String, DataInstance> sentsf = data.next().first();
for(Entry<String, DataInstance> en: sentsf.entrySet()){
for(CoreLabel l : en.getValue().getTokens()){
for(Entry<String, Class<? extends TypesafeMap.Key<String>>> enc: constVars.getAnswerClass().entrySet()){
if(l.get(enc.getValue()).equals(enc.getKey())){
counter.incrementCount(enc.getKey());
}
}
}
}
}
return counter;
}
Counter<CandidatePhrase> closeToPositivesFirstIter = null;
Counter<CandidatePhrase> closeToNegativesFirstIter = null;
public class ChooseDatumsThread implements Callable {
Collection<String> keys;
Map<String, DataInstance> sents;
Class answerClass;
String answerLabel;
TwoDimensionalCounter<CandidatePhrase, E> wordsPatExtracted;
Counter<E> allSelectedPatterns;
Counter<Integer> wordClassClustersOfPositive;
Map<String, Collection<CandidatePhrase>> allPossiblePhrases;
boolean expandPos;
boolean expandNeg;
public ChooseDatumsThread(String label, Map<String, DataInstance> sents, Collection<String> keys, TwoDimensionalCounter<CandidatePhrase, E> wordsPatExtracted, Counter<E> allSelectedPatterns,
Counter<Integer> wordClassClustersOfPositive, Map<String, Collection<CandidatePhrase>> allPossiblePhrases, boolean expandPos, boolean expandNeg){
this.answerLabel = label;
this.sents = sents;
this.keys = keys;
this.wordsPatExtracted = wordsPatExtracted;
this.allSelectedPatterns = allSelectedPatterns;
this.wordClassClustersOfPositive = wordClassClustersOfPositive;
this.allPossiblePhrases = allPossiblePhrases;
answerClass = constVars.getAnswerClass().get(answerLabel);
this.expandNeg = expandNeg;
this.expandPos = expandPos;
}
@Override
public Quintuple<Set<CandidatePhrase>, Set<CandidatePhrase>, Set<CandidatePhrase>, Counter<CandidatePhrase>, Counter<CandidatePhrase>> call() throws Exception {
Random r = new Random(10);
Random rneg = new Random(10);
Set<CandidatePhrase> allPositivePhrases = new HashSet<>();
Set<CandidatePhrase> allNegativePhrases = new HashSet<>();
Set<CandidatePhrase> allUnknownPhrases = new HashSet<>();
Counter<CandidatePhrase> allCloseToPositivePhrases = new ClassicCounter<>();
Counter<CandidatePhrase> allCloseToNegativePhrases = new ClassicCounter<>();
Set<CandidatePhrase> knownPositivePhrases = CollectionUtils.unionAsSet(constVars.getLearnedWords(answerLabel).keySet(), constVars.getSeedLabelDictionary().get(answerLabel));
Set<CandidatePhrase> allConsideredPhrases = new HashSet<>();
Map<Class, Object> otherIgnoreClasses = constVars.getIgnoreWordswithClassesDuringSelection().get(answerLabel);
int numlabeled = 0;
for (String sentid : keys) {
DataInstance sentInst = sents.get(sentid);
List<CoreLabel> value = sentInst.getTokens();
CoreLabel[] sent = value.toArray(new CoreLabel[value.size()]);
for (int i = 0; i < sent.length; i++) {
CoreLabel l = sent[i];
if (l.get(answerClass).equals(answerLabel)) {
numlabeled++;
CandidatePhrase candidate = l.get(PatternsAnnotations.LongestMatchedPhraseForEachLabel.class).get(answerLabel);
if (candidate == null) {
throw new RuntimeException("for sentence id " + sentid + " and token id " + i + " candidate is null for " + l.word() + " and longest matching" + l.get(PatternsAnnotations.LongestMatchedPhraseForEachLabel.class) + " and matched phrases are " + l.get(PatternsAnnotations.MatchedPhrases.class));
//candidate = CandidatePhrase.createOrGet(l.word());
}
//If the phrase does not exist in its form in the datset (happens when fuzzy matching etc).
if(!Data.rawFreq.containsKey(candidate)){
candidate = CandidatePhrase.createOrGet(l.word());
}
//Do not add to positive if the word is a "negative" (stop word, english word, ...)
if(hasElement(allPossiblePhrases, candidate, answerLabel) || PatternFactory.ignoreWordRegex.matcher(candidate.getPhrase()).matches())
continue;
allPositivePhrases.add(candidate);
} else {
Map<String, CandidatePhrase> longestMatching = l.get(PatternsAnnotations.LongestMatchedPhraseForEachLabel.class);
boolean ignoreclass = false;
CandidatePhrase candidate = CandidatePhrase.createOrGet(l.word());
for (Class cl : otherIgnoreClasses.keySet()) {
if ((Boolean) l.get(cl)) {
ignoreclass = true;
candidate = longestMatching.containsKey("OTHERSEM")? longestMatching.get("OTHERSEM") : candidate;
break;
}
}
if(!ignoreclass) {
ignoreclass = constVars.functionWords.contains(l.word());
}
boolean negative = false;
boolean add= false;
for (Map.Entry<String, CandidatePhrase> lo : longestMatching.entrySet()) {
//assert !lo.getValue().getPhrase().isEmpty() : "How is the longestmatching phrase for " + l.word() + " empty ";
if (!lo.getKey().equals(answerLabel) && lo.getValue() != null) {
negative = true;
add = true;
//If the phrase does not exist in its form in the datset (happens when fuzzy matching etc).
if(Data.rawFreq.containsKey(lo.getValue())){
candidate = lo.getValue();
}
}
}
if (!negative && ignoreclass) {
add = true;
}
if(add && rneg.nextDouble() < constVars.perSelectNeg){
assert !candidate.getPhrase().isEmpty();
allNegativePhrases.add(candidate);
}
if(!negative && !ignoreclass && (expandPos || expandNeg) && !hasElement(allPossiblePhrases, candidate, answerLabel) && !PatternFactory.ignoreWordRegex.matcher(candidate.getPhrase()).matches()) {
if (!allConsideredPhrases.contains(candidate)) {
Pair<Counter<CandidatePhrase>, Counter<CandidatePhrase>> sims;
assert candidate != null;
if(constVars.useWordVectorsToComputeSim)
sims = computeSimWithWordVectors(Arrays.asList(candidate), knownPositivePhrases, allPossiblePhrases, answerLabel);
else
sims = computeSimWithWordCluster(Arrays.asList(candidate), knownPositivePhrases, new AtomicDouble());
boolean addedAsPos = false;
if(expandPos)
{
double sim = sims.first().getCount(candidate);
if (sim > constVars.similarityThresholdHighPrecision){
allCloseToPositivePhrases.setCount(candidate, sim);
addedAsPos = true;
}
}
if(expandNeg && !addedAsPos) {
double simneg = sims.second().getCount(candidate);
if (simneg > constVars.similarityThresholdHighPrecision)
allCloseToNegativePhrases.setCount(candidate, simneg);
}
allConsideredPhrases.add(candidate);
}
}
}
}
allUnknownPhrases.addAll(chooseUnknownPhrases(sentInst, r, constVars.perSelectRand, constVars.getAnswerClass().get(answerLabel), answerLabel, Math.max(0, Integer.MAX_VALUE)));
//
// if (negative && getRandomBoolean(rneg, perSelectNeg)) {
// numneg++;
// } else if (getRandomBoolean(r, perSelectRand)) {
// candidate = CandidatePhrase.createOrGet(l.word());
// numneg++;
// } else {
// continue;
// }
//
//
// chosen.add(new Pair<String, Integer>(en.getKey(), i));
}
return new Quintuple(allPositivePhrases, allNegativePhrases, allUnknownPhrases, allCloseToPositivePhrases, allCloseToNegativePhrases);
}
}
static private class PhrasePair{
final String p1;
final String p2;
final int hashCode;
public PhrasePair(String p1, String p2) {
if(p1.compareTo(p2) <=0)
{
this.p1 = p1;
this.p2 = p2;
}else
{
this.p1 = p2;
this.p2 = p1;
}
this.hashCode = p1.hashCode() + p2.hashCode() + 331;
}
@Override
public int hashCode(){
return hashCode;
}
@Override
public boolean equals(Object o) {
if (!(o instanceof PhrasePair))
return false;
PhrasePair p = (PhrasePair) o;
if (p.getPhrase1().equals(this.getPhrase1()) && p.getPhrase2().equals(this.getPhrase2()))
return true;
return false;
}
public String getPhrase1() {
return p1;
}
public String getPhrase2() {
return p2;
}
}
static Counter<PhrasePair> cacheSimilarities = new ConcurrentHashCounter<>();
//First map is phrase, second map is label to similarity stats
static Map<String, Map<String, double[]>> similaritiesWithLabeledPhrases = new ConcurrentHashMap<>();
Map<String, Collection<CandidatePhrase>> getAllPossibleNegativePhrases(String answerLabel){
//make all possible negative phrases
Map<String, Collection<CandidatePhrase>> allPossiblePhrases = new HashMap<>();
Collection<CandidatePhrase> negPhrases = new HashSet<>();
//negPhrases.addAll(constVars.getOtherSemanticClassesWords());
negPhrases.addAll(constVars.getStopWords());
negPhrases.addAll(CandidatePhrase.convertStringPhrases(constVars.functionWords));
negPhrases.addAll(CandidatePhrase.convertStringPhrases(constVars.getEnglishWords()));
allPossiblePhrases.put("NEGATIVE", negPhrases);
for(String label: constVars.getLabels()) {
if (!label.equals(answerLabel)){
allPossiblePhrases.put(label, new HashSet<>());
if(constVars.getLearnedWordsEachIter().containsKey(label))
allPossiblePhrases.get(label).addAll(constVars.getLearnedWords(label).keySet());
allPossiblePhrases.get(label).addAll(constVars.getSeedLabelDictionary().get(label));
}
}
allPossiblePhrases.put("OTHERSEM", constVars.getOtherSemanticClassesWords());
return allPossiblePhrases;
}
public GeneralDataset<String, ScorePhraseMeasures> choosedatums(boolean forLearningPattern, String answerLabel,
TwoDimensionalCounter<CandidatePhrase, E> wordsPatExtracted,
Counter<E> allSelectedPatterns, boolean computeRawFreq) throws IOException {
boolean expandNeg = false;
if(closeToNegativesFirstIter == null){
closeToNegativesFirstIter = new ClassicCounter<>();
if(constVars.expandNegativesWhenSampling)
expandNeg = true;
}
boolean expandPos = false;
if(closeToPositivesFirstIter == null) {
closeToPositivesFirstIter = new ClassicCounter<>();
if(constVars.expandPositivesWhenSampling)
expandPos = true;
}
Counter<Integer> distSimClustersOfPositive = new ClassicCounter<>();
if((expandPos || expandNeg) && !constVars.useWordVectorsToComputeSim){
for(CandidatePhrase s: CollectionUtils.union(constVars.getLearnedWords(answerLabel).keySet(), constVars.getSeedLabelDictionary().get(answerLabel))){
String[] toks = s.getPhrase().split("\\s+");
Integer num = constVars.getWordClassClusters().get(s.getPhrase());
if(num == null)
num = constVars.getWordClassClusters().get(s.getPhrase().toLowerCase());
if(num == null){
for(String tok: toks){
Integer toknum =constVars.getWordClassClusters().get(tok);
if(toknum == null)
toknum =constVars.getWordClassClusters().get(tok.toLowerCase());
if(toknum != null){
distSimClustersOfPositive.incrementCount(toknum);
}
}
} else
distSimClustersOfPositive.incrementCount(num);
}
}
//computing this regardless of expandpos and expandneg because we reject all positive words that occur in negatives (can happen in multi word phrases etc)
Map<String, Collection<CandidatePhrase>> allPossibleNegativePhrases = getAllPossibleNegativePhrases(answerLabel);
GeneralDataset<String, ScorePhraseMeasures> dataset = new RVFDataset<>();
int numpos = 0;
Set<CandidatePhrase> allNegativePhrases = new HashSet<>();
Set<CandidatePhrase> allUnknownPhrases = new HashSet<>();
Set<CandidatePhrase> allPositivePhrases = new HashSet<>();
//Counter<CandidatePhrase> allCloseToPositivePhrases = new ClassicCounter<CandidatePhrase>();
//Counter<CandidatePhrase> allCloseToNegativePhrases = new ClassicCounter<CandidatePhrase>();
//for all sentences brtch
ConstantsAndVariables.DataSentsIterator sentsIter = new ConstantsAndVariables.DataSentsIterator(constVars.batchProcessSents);
while(sentsIter.hasNext()) {
Pair<Map<String, DataInstance>, File> sentsf = sentsIter.next();
Map<String, DataInstance> sents = sentsf.first();
Redwood.log(Redwood.DBG, "Sampling datums from " + sentsf.second());
if (computeRawFreq)
Data.computeRawFreqIfNull(sents, PatternFactory.numWordsCompoundMax);
List<List<String>> threadedSentIds = GetPatternsFromDataMultiClass.getThreadBatches(new ArrayList<>(sents.keySet()), constVars.numThreads);
ExecutorService executor = Executors.newFixedThreadPool(constVars.numThreads);
List<Future<Quintuple<Set<CandidatePhrase>, Set<CandidatePhrase>, Set<CandidatePhrase>, Counter<CandidatePhrase>, Counter<CandidatePhrase>>>> list = new ArrayList<>();
//multi-threaded choose positive, negative and unknown
for (List<String> keys : threadedSentIds) {
Callable<Quintuple<Set<CandidatePhrase>, Set<CandidatePhrase>, Set<CandidatePhrase>, Counter<CandidatePhrase>, Counter<CandidatePhrase>>> task = new ChooseDatumsThread(answerLabel, sents, keys,
wordsPatExtracted, allSelectedPatterns, distSimClustersOfPositive, allPossibleNegativePhrases, expandPos, expandNeg);
Future<Quintuple<Set<CandidatePhrase>, Set<CandidatePhrase>, Set<CandidatePhrase>, Counter<CandidatePhrase>, Counter<CandidatePhrase>>> submit = executor.submit(task);
list.add(submit);
}
// Now retrieve the result
for (Future<Quintuple<Set<CandidatePhrase>, Set<CandidatePhrase>, Set<CandidatePhrase>, Counter<CandidatePhrase>, Counter<CandidatePhrase>>> future : list) {
try {
Quintuple<Set<CandidatePhrase>, Set<CandidatePhrase>, Set<CandidatePhrase>, Counter<CandidatePhrase>, Counter<CandidatePhrase>> result = future.get();
allPositivePhrases.addAll(result.first());
allNegativePhrases.addAll(result.second());
allUnknownPhrases.addAll(result.third());
if(expandPos)
for(Entry<CandidatePhrase, Double> en : result.fourth().entrySet())
closeToPositivesFirstIter.setCount(en.getKey(), en.getValue());
if(expandNeg)
for(Entry<CandidatePhrase, Double> en : result.fifth().entrySet())
closeToNegativesFirstIter.setCount(en.getKey(), en.getValue());
} catch (Exception e) {
executor.shutdownNow();
throw new RuntimeException(e);
}
}
executor.shutdown();
}
//Set<CandidatePhrase> knownPositivePhrases = CollectionUtils.unionAsSet(constVars.getLearnedWords().get(answerLabel).keySet(), constVars.getSeedLabelDictionary().get(answerLabel));
//TODO: this is kinda not nice; how is allpositivephrases different from positivephrases again?
allPositivePhrases.addAll(constVars.getLearnedWords(answerLabel).keySet());
//allPositivePhrases.addAll(knownPositivePhrases);
BufferedWriter logFile = null;
BufferedWriter logFileFeat = null;
if(constVars.logFileVectorSimilarity != null){
logFile = new BufferedWriter(new FileWriter(constVars.logFileVectorSimilarity));
logFileFeat = new BufferedWriter(new FileWriter(constVars.logFileVectorSimilarity+"_feat"));
if(wordVectors != null){
for(CandidatePhrase p : allPositivePhrases){
if(wordVectors.containsKey(p.getPhrase())){
logFile.write(p.getPhrase()+"-P " + ArrayUtils.toString(wordVectors.get(p.getPhrase()), " ")+"\n");
}
}
}
}
if(constVars.expandPositivesWhenSampling){
//TODO: patwtbyfrew
//Counters.retainTop(allCloseToPositivePhrases, (int) (allCloseToPositivePhrases.size()*constVars.subSampleUnkAsPosUsingSimPercentage));
Redwood.log("Expanding positives by adding " + Counters.toSortedString(closeToPositivesFirstIter, closeToPositivesFirstIter.size(),"%1$s:%2$f", "\t")+ " phrases");
allPositivePhrases.addAll(closeToPositivesFirstIter.keySet());
//write log
if(logFile != null && wordVectors != null && expandNeg){
for(CandidatePhrase p : closeToPositivesFirstIter.keySet()){
if(wordVectors.containsKey(p.getPhrase())){
logFile.write(p.getPhrase()+"-PP " + ArrayUtils.toString(wordVectors.get(p.getPhrase()), " ")+"\n");
}
}
}
}
if(constVars.expandNegativesWhenSampling){
//TODO: patwtbyfrew
//Counters.retainTop(allCloseToPositivePhrases, (int) (allCloseToPositivePhrases.size()*constVars.subSampleUnkAsPosUsingSimPercentage));
Redwood.log("Expanding negatives by adding " + Counters.toSortedString(closeToNegativesFirstIter , closeToNegativesFirstIter.size(), "%1$s:%2$f","\t")+ " phrases");
allNegativePhrases.addAll(closeToNegativesFirstIter.keySet());
//write log
if(logFile != null && wordVectors != null && expandNeg){
for(CandidatePhrase p : closeToNegativesFirstIter.keySet()){
if(wordVectors.containsKey(p.getPhrase())){
logFile.write(p.getPhrase()+"-NN " + ArrayUtils.toString(wordVectors.get(p.getPhrase()), " ")+"\n");
}
}
}
}
System.out.println("all positive phrases of size " + allPositivePhrases.size() + " are " + allPositivePhrases);
for(CandidatePhrase candidate: allPositivePhrases) {
Counter<ScorePhraseMeasures> feat;