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
|
#!/usr/bin/perl
###############################################
# An implementation of paired bootstrap resampling for testing the statistical
# significance of the difference between two systems from (Koehn 2004 @ EMNLP)
#
# Usage: ./compare-hypotheses-with-significance.pl hypothesis_1 hypothesis_2 reference_1 [ reference_2 ... ]
#
# Author: Mark Fishel, fishel@ut.ee
#
# 22.10: altered algorithm according to (Riezler and Maxwell 2005 @ MTSE'05), now computes p-value
###############################################
use strict;
#constants
my $TIMES_TO_REPEAT_SUBSAMPLING = 1000;
my $SUBSAMPLE_SIZE = 0; # if 0 then subsample size is equal to the whole set
my $MAX_NGRAMS_FOR_BLEU = 4;
#checking cmdline argument consistency
if (@ARGV < 3) {
print STDERR "Usage: ./bootstrap-hypothesis-difference-significance.pl hypothesis_1 hypothesis_2 reference_1 [ reference_2 ... ]\n";
unless ($ARGV[0] =~ /^(--help|-help|-h|-\?|\/\?|--usage|-usage)$/) {
die("\nERROR: not enough arguments");
}
exit 1;
}
print "reading data; " . `date`;
#read all data
my $data = readAllData(@ARGV);
#calculate each sentence's contribution to BP and ngram precision
print "performing preliminary calculations (hypothesis 1); " . `date`;
preEvalHypo($data, "hyp1");
print "performing preliminary calculations (hypothesis 2); " . `date`;
preEvalHypo($data, "hyp2");
#start comparing
print "comparing hypotheses; " . `date`;
my @subSampleBleuDiffArr;
my @subSampleBleu1Arr;
my @subSampleBleu2Arr;
#applying sampling
for (1..$TIMES_TO_REPEAT_SUBSAMPLING) {
my $subSampleIndices = drawWithReplacement($data->{size}, ($SUBSAMPLE_SIZE? $SUBSAMPLE_SIZE: $data->{size}));
my $bleu1 = getBleu($data->{refs}, $data->{hyp1}, $subSampleIndices);
my $bleu2 = getBleu($data->{refs}, $data->{hyp2}, $subSampleIndices);
push @subSampleBleuDiffArr, abs($bleu2 - $bleu1);
push @subSampleBleu1Arr, $bleu1;
push @subSampleBleu2Arr, $bleu2;
if ($_ % int($TIMES_TO_REPEAT_SUBSAMPLING / 100) == 0) {
print "$_ / $TIMES_TO_REPEAT_SUBSAMPLING " . `date`;
}
}
#get subsample bleu difference mean
my $averageSubSampleBleuDiff = 0;
for my $subSampleDiff (@subSampleBleuDiffArr) {
$averageSubSampleBleuDiff += $subSampleDiff;
}
$averageSubSampleBleuDiff /= $TIMES_TO_REPEAT_SUBSAMPLING;
print "average subsample bleu: $averageSubSampleBleuDiff " . `date`;
#calculating p-value
my $count = 0;
my $realBleu1 = getBleu($data->{refs}, $data->{hyp1});
my $realBleu2 = getBleu($data->{refs}, $data->{hyp2});
print "actual BLEU of hypothesis 1: $realBleu1\n";
print "actual BLEU of hypothesis 1: $realBleu2\n";
my $realBleuDiff = abs($realBleu2 - $realBleu1);
for my $subSampleDiff (@subSampleBleuDiffArr) {
my $op;
if ($subSampleDiff - $averageSubSampleBleuDiff >= $realBleuDiff) {
$count++;
$op = ">=";
}
else {
$op = "< ";
}
#print "$subSampleDiff - $averageSubSampleBleuDiff $op $realBleuDiff\n";
}
my $result = $count / $TIMES_TO_REPEAT_SUBSAMPLING;
print "Assuming that essentially the same system generated the two hypothesis translations (null-hypothesis),\n";
print "the probability of actually getting them (p-value) is: $result.\n";
my @sorted1 = sort @subSampleBleu1Arr;
my @sorted2 = sort @subSampleBleu2Arr;
print "95% confidence interval for hypothesis 1: " . $sorted1[25] . " -- " . $sorted1[924] . "\n";
print "95% confidence interval for hypothesis 2: " . $sorted2[25] . " -- " . $sorted2[924] . "\n";
#####
# read 2 hyp and 1 to \infty ref data files
#####
sub readAllData {
my ($hypFile1, $hypFile2, @refFiles) = @_;
my %result;
#reading hypotheses and checking for matching sizes
$result{hyp1} = readData($hypFile1);
$result{size} = scalar @{$result{hyp1}};
$result{hyp2} = readData($hypFile2);
unless (scalar @{$result{hyp2}} == $result{size}) {
die ("ERROR: sizes of hypothesis sets 1 and 2 don't match");
}
#reading reference(s) and checking for matching sizes
$result{refs} = [];
my $i = 0;
for my $refFile (@refFiles) {
$i++;
my $refDataX = readData($refFile);
unless (scalar @$refDataX == $result{size}) {
die ("ERROR: ref set $i size doesn't match the size of hyp sets");
}
push @{$result{refs}}, $refDataX;
}
return \%result;
}
#####
# read sentences from file
#####
sub readData {
my $file = shift;
my @result;
open (FILE, $file) or die ("Failed to open `$file' for reading");
while (<FILE>) {
push @result, { words => [split(/\s+/, $_)] };
}
close (FILE);
return \@result;
}
#####
# calculate each sentence's contribution to the ngram precision and brevity penalty
#####
sub preEvalHypo {
my $data = shift;
my $hypId = shift;
my ($correctNgramCounts, $totalNgramCounts);
my ($refNgramCounts, $hypNgramCounts);
for my $lineIdx (0..($data->{size} - 1)) {
my $hypSnt = $data->{$hypId}->[$lineIdx];
#update total hyp len
$hypSnt->{hyplen} = scalar @{$hypSnt->{words}};
#update total ref len with closest current ref len
$hypSnt->{reflen} = getClosestLength($data->{refs}, $lineIdx, $hypSnt->{hyplen});
$hypSnt->{correctNgrams} = [];
$hypSnt->{totalNgrams} = [];
#update ngram precision for each n-gram order
for my $order (1..$MAX_NGRAMS_FOR_BLEU) {
#hyp ngrams
$hypNgramCounts = groupNgrams($hypSnt, $order);
#ref ngrams
$refNgramCounts = groupNgramsMultiSrc($data->{refs}, $lineIdx, $order);
$correctNgramCounts = 0;
$totalNgramCounts = 0;
#correct, total
for my $ngram (keys %$hypNgramCounts) {
$correctNgramCounts += min($hypNgramCounts->{$ngram}, $refNgramCounts->{$ngram});
$totalNgramCounts += $hypNgramCounts->{$ngram};
}
$hypSnt->{correctNgrams}->[$order] = $correctNgramCounts;
$hypSnt->{totalNgrams}->[$order] = $totalNgramCounts;
}
}
}
#####
# draw a subsample of size $subSize from set (0..$setSize) with replacement
#####
sub drawWithReplacement {
my ($setSize, $subSize) = @_;
my @result;
for (1..$subSize) {
push @result, int(rand($setSize));
}
return \@result;
}
#####
# refs: arrayref of different references, reference = array of lines, line = array of words, word = string
# hyp: arrayref of lines of hypothesis translation, line = array of words, word = string
# idxs: indices of lines to include; default value - full set (0..size_of_hyp-1)
#####
sub getBleu {
my ($refs, $hyp, $idxs) = @_;
#default value for $idxs
unless (defined($idxs)) {
$idxs = [0..((scalar @$hyp) - 1)];
}
#vars
my ($hypothesisLength, $referenceLength) = (0, 0);
my (@correctNgramCounts, @totalNgramCounts);
my ($refNgramCounts, $hypNgramCounts);
#gather info from each line
for my $lineIdx (@$idxs) {
my $hypSnt = $hyp->[$lineIdx];
#update total hyp len
$hypothesisLength += $hypSnt->{hyplen};
#update total ref len with closest current ref len
$referenceLength += $hypSnt->{reflen};
#update ngram precision for each n-gram order
for my $order (1..$MAX_NGRAMS_FOR_BLEU) {
$correctNgramCounts[$order] += $hypSnt->{correctNgrams}->[$order];
$totalNgramCounts[$order] += $hypSnt->{totalNgrams}->[$order];
}
}
#compose bleu score
my $brevityPenalty = ($hypothesisLength < $referenceLength)? exp(1 - $referenceLength/$hypothesisLength): 1;
my $logsum = 0;
for my $order (1..$MAX_NGRAMS_FOR_BLEU) {
$logsum += safeLog($correctNgramCounts[$order] / $totalNgramCounts[$order]);
}
return $brevityPenalty * exp($logsum / $MAX_NGRAMS_FOR_BLEU);
}
#####
#
#####
sub getClosestLength {
my ($refs, $lineIdx, $hypothesisLength) = @_;
my $bestDiff = infty();
my $bestLen = infty();
my ($currLen, $currDiff);
for my $ref (@$refs) {
$currLen = scalar @{$ref->[$lineIdx]->{words}};
$currDiff = abs($currLen - $hypothesisLength);
if ($currDiff < $bestDiff or ($currDiff == $bestDiff and $currLen < $bestLen)) {
$bestDiff = $currDiff;
$bestLen = $currLen;
}
}
return $bestLen;
}
#####
#
#####
sub groupNgrams {
my ($snt, $order) = @_;
my %result;
my $size = scalar @{$snt->{words}};
my $ngram;
for my $i (0..($size-$order)) {
$ngram = join(" ", @{$snt->{words}}[$i..($i + $order - 1)]);
$result{$ngram}++;
}
return \%result;
}
#####
#
#####
sub groupNgramsMultiSrc {
my ($refs, $lineIdx, $order) = @_;
my %result;
for my $ref (@$refs) {
my $currNgramCounts = groupNgrams($ref->[$lineIdx], $order);
for my $currNgram (keys %$currNgramCounts) {
$result{$currNgram} = max($result{$currNgram}, $currNgramCounts->{$currNgram});
}
}
return \%result;
}
#####
#
#####
sub safeLog {
my $x = shift;
return ($x > 0)? log($x): -infty();
}
#####
#
#####
sub infty {
return 1e6000;
}
#####
#
#####
sub min {
my ($a, $b) = @_;
return ($a < $b)? $a: $b;
}
#####
#
#####
sub max {
my ($a, $b) = @_;
return ($a > $b)? $a: $b;
}
|