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
|
#include "BleuScorer.h"
#include <algorithm>
#include <cassert>
#include <cmath>
#include <climits>
#include <fstream>
#include <iostream>
#include <stdexcept>
#include "util/check.hh"
#include "Ngram.h"
#include "Reference.h"
#include "Util.h"
#include "Vocabulary.h"
using namespace std;
namespace {
// configure regularisation
const char KEY_REFLEN[] = "reflen";
const char REFLEN_AVERAGE[] = "average";
const char REFLEN_SHORTEST[] = "shortest";
const char REFLEN_CLOSEST[] = "closest";
} // namespace
BleuScorer::BleuScorer(const string& config)
: StatisticsBasedScorer("BLEU", config),
m_ref_length_type(CLOSEST) {
const string reflen = getConfig(KEY_REFLEN, REFLEN_CLOSEST);
if (reflen == REFLEN_AVERAGE) {
m_ref_length_type = AVERAGE;
} else if (reflen == REFLEN_SHORTEST) {
m_ref_length_type = SHORTEST;
} else if (reflen == REFLEN_CLOSEST) {
m_ref_length_type = CLOSEST;
} else {
throw runtime_error("Unknown reference length strategy: " + reflen);
}
}
BleuScorer::~BleuScorer() {}
size_t BleuScorer::CountNgrams(const string& line, NgramCounts& counts,
unsigned int n)
{
assert(n > 0);
vector<int> encoded_tokens;
TokenizeAndEncode(line, encoded_tokens);
for (size_t k = 1; k <= n; ++k) {
//ngram order longer than sentence - no point
if (k > encoded_tokens.size()) {
continue;
}
for (size_t i = 0; i < encoded_tokens.size()-k+1; ++i) {
vector<int> ngram;
for (size_t j = i; j < i+k && j < encoded_tokens.size(); ++j) {
ngram.push_back(encoded_tokens[j]);
}
counts.Add(ngram);
}
}
return encoded_tokens.size();
}
void BleuScorer::setReferenceFiles(const vector<string>& referenceFiles)
{
// Make sure reference data is clear
m_references.reset();
mert::VocabularyFactory::GetVocabulary()->clear();
//load reference data
for (size_t i = 0; i < referenceFiles.size(); ++i) {
TRACE_ERR("Loading reference from " << referenceFiles[i] << endl);
if (!OpenReference(referenceFiles[i].c_str(), i)) {
throw runtime_error("Unable to open " + referenceFiles[i]);
}
}
}
bool BleuScorer::OpenReference(const char* filename, size_t file_id) {
ifstream ifs(filename);
if (!ifs) {
cerr << "Cannot open " << filename << endl;
return false;
}
return OpenReferenceStream(&ifs, file_id);
}
bool BleuScorer::OpenReferenceStream(istream* is, size_t file_id) {
if (is == NULL) return false;
string line;
size_t sid = 0;
while (getline(*is, line)) {
line = preprocessSentence(line);
if (file_id == 0) {
Reference* ref = new Reference;
m_references.push_back(ref); // Take ownership of the Reference object.
}
if (m_references.size() <= sid) {
cerr << "Reference " << file_id << "has too many sentences." << endl;
return false;
}
NgramCounts counts;
size_t length = CountNgrams(line, counts, kBleuNgramOrder);
//for any counts larger than those already there, merge them in
for (NgramCounts::const_iterator ci = counts.begin(); ci != counts.end(); ++ci) {
const NgramCounts::Key& ngram = ci->first;
const NgramCounts::Value newcount = ci->second;
NgramCounts::Value oldcount = 0;
m_references[sid]->get_counts()->Lookup(ngram, &oldcount);
if (newcount > oldcount) {
m_references[sid]->get_counts()->operator[](ngram) = newcount;
}
}
//add in the length
m_references[sid]->push_back(length);
if (sid > 0 && sid % 100 == 0) {
TRACE_ERR(".");
}
++sid;
}
return true;
}
void BleuScorer::prepareStats(size_t sid, const string& text, ScoreStats& entry)
{
if (sid >= m_references.size()) {
stringstream msg;
msg << "Sentence id (" << sid << ") not found in reference set";
throw runtime_error(msg.str());
}
NgramCounts testcounts;
// stats for this line
vector<ScoreStatsType> stats(kBleuNgramOrder * 2);
string sentence = preprocessSentence(text);
const size_t length = CountNgrams(sentence, testcounts, kBleuNgramOrder);
const int reference_len = CalcReferenceLength(sid, length);
stats.push_back(reference_len);
//precision on each ngram type
for (NgramCounts::const_iterator testcounts_it = testcounts.begin();
testcounts_it != testcounts.end(); ++testcounts_it) {
const NgramCounts::Value guess = testcounts_it->second;
const size_t len = testcounts_it->first.size();
NgramCounts::Value correct = 0;
NgramCounts::Value v = 0;
if (m_references[sid]->get_counts()->Lookup(testcounts_it->first, &v)) {
correct = min(v, guess);
}
stats[len * 2 - 2] += correct;
stats[len * 2 - 1] += guess;
}
entry.set(stats);
}
statscore_t BleuScorer::calculateScore(const vector<int>& comps) const
{
CHECK(comps.size() == kBleuNgramOrder * 2 + 1);
float logbleu = 0.0;
for (int i = 0; i < kBleuNgramOrder; ++i) {
if (comps[2*i] == 0) {
return 0.0;
}
logbleu += log(comps[2*i]) - log(comps[2*i+1]);
}
logbleu /= kBleuNgramOrder;
// reflength divided by test length
const float brevity = 1.0 - static_cast<float>(comps[kBleuNgramOrder * 2]) / comps[1];
if (brevity < 0.0) {
logbleu += brevity;
}
return exp(logbleu);
}
int BleuScorer::CalcReferenceLength(size_t sentence_id, size_t length) {
switch (m_ref_length_type) {
case AVERAGE:
return m_references[sentence_id]->CalcAverage();
break;
case CLOSEST:
return m_references[sentence_id]->CalcClosest(length);
break;
case SHORTEST:
return m_references[sentence_id]->CalcShortest();
break;
default:
cerr << "unknown reference types." << endl;
exit(1);
}
}
void BleuScorer::DumpCounts(ostream* os,
const NgramCounts& counts) const {
for (NgramCounts::const_iterator it = counts.begin();
it != counts.end(); ++it) {
*os << "(";
const NgramCounts::Key& keys = it->first;
for (size_t i = 0; i < keys.size(); ++i) {
if (i != 0) {
*os << " ";
}
*os << keys[i];
}
*os << ") : " << it->second << ", ";
}
*os << endl;
}
float sentenceLevelBleuPlusOne(const vector<float>& stats) {
CHECK(stats.size() == kBleuNgramOrder * 2 + 1);
float logbleu = 0.0;
for (int j = 0; j < kBleuNgramOrder; j++) {
logbleu += log(stats[2 * j] + 1.0) - log(stats[2 * j + 1] + 1.0);
}
logbleu /= kBleuNgramOrder;
const float brevity = 1.0 - stats[(kBleuNgramOrder * 2)] / stats[1];
if (brevity < 0.0) {
logbleu += brevity;
}
return exp(logbleu);
}
float sentenceLevelBackgroundBleu(const std::vector<float>& sent, const std::vector<float>& bg)
{
// Sum sent and background
std::vector<float> stats;
CHECK(sent.size()==bg.size());
CHECK(sent.size()==kBleuNgramOrder*2+1);
for(size_t i=0;i<sent.size();i++)
stats.push_back(sent[i]+bg[i]);
// Calculate BLEU
float logbleu = 0.0;
for (int j = 0; j < kBleuNgramOrder; j++) {
logbleu += log(stats[2 * j]) - log(stats[2 * j + 1]);
}
logbleu /= kBleuNgramOrder;
const float brevity = 1.0 - stats[(kBleuNgramOrder * 2)] / stats[1];
if (brevity < 0.0) {
logbleu += brevity;
}
// Exponentiate and scale by reference length (as per Chiang et al 08)
return exp(logbleu) * stats[kBleuNgramOrder*2];
}
float unsmoothedBleu(const std::vector<float>& stats) {
CHECK(stats.size() == kBleuNgramOrder * 2 + 1);
float logbleu = 0.0;
for (int j = 0; j < kBleuNgramOrder; j++) {
logbleu += log(stats[2 * j]) - log(stats[2 * j + 1]);
}
logbleu /= kBleuNgramOrder;
const float brevity = 1.0 - stats[(kBleuNgramOrder * 2)] / stats[1];
if (brevity < 0.0) {
logbleu += brevity;
}
return exp(logbleu);
}
|