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
|
#include "BleuDocScorer.h"
#include <sys/types.h>
#include <algorithm>
#include <cassert>
#include <cmath>
#include <climits>
#include <fstream>
#include <iostream>
#include <stdexcept>
#include "util/exception.hh"
#include "Ngram.h"
#include "Reference.h"
#include "Util.h"
#include "Vocabulary.h"
using namespace std;
#if defined __MINGW32__
#ifndef uint
#define uint uint16_t
#endif // uint
#endif // if
namespace
{
// configure regularisation
const char KEY_REFLEN[] = "reflen";
const char REFLEN_AVERAGE[] = "average";
const char REFLEN_SHORTEST[] = "shortest";
const char REFLEN_CLOSEST[] = "closest";
} // namespace
namespace MosesTuning
{
BleuDocScorer::BleuDocScorer(const string& config)
: BleuScorer("BLEUDOC", 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);
}
}
BleuDocScorer::~BleuDocScorer() {}
bool BleuDocScorer::OpenReferenceStream(istream* is, size_t file_id)
{
if (is == NULL) return false;
string line;
size_t doc_id = -1;
size_t sid = 0;
while (getline(*is, line)) {
if (line.find("<doc docid") != std::string::npos) { // new document
doc_id++;
m_references.push_back(new ScopedVector<Reference>());
sid = 0;
} else if (line.find("<seg") != std::string::npos) { //new sentence
int start = line.find_first_of('>') + 1;
std::string trans = line.substr(start, line.find_last_of('<')-start);
trans = preprocessSentence(trans);
if (file_id == 0) {
Reference* ref = new Reference;
m_references[doc_id]->push_back(ref); // Take ownership of the Reference object.
}
if (m_references[doc_id]->size() <= sid) {
return false;
}
NgramCounts counts;
size_t length = CountNgrams(trans, 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[doc_id]->get().at(sid)->get_counts()->Lookup(ngram, &oldcount);
if (newcount > oldcount) {
m_references[doc_id]->get().at(sid)->get_counts()->operator[](ngram) = newcount;
}
}
//add in the length
m_references[doc_id]->get().at(sid)->push_back(length);
if (sid > 0 && sid % 100 == 0) {
TRACE_ERR(".");
}
++sid;
}
}
return true;
}
void BleuDocScorer::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());
}
std::vector<std::string> sentences = splitDoc(text);
vector<ScoreStatsType> totStats(kBleuNgramOrder * 2 + 1);
for (uint i=0; i<sentences.size(); ++i) {
NgramCounts testcounts;
// stats for this line
vector<ScoreStatsType> stats(kBleuNgramOrder * 2);
string sentence = preprocessSentence(sentences[i]);
const size_t length = CountNgrams(sentence, testcounts, kBleuNgramOrder);
//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().at(i)->get_counts()->Lookup(testcounts_it->first, &v)) {
correct = min(v, guess);
}
stats[len * 2 - 2] += correct;
stats[len * 2 - 1] += guess;
}
const int reference_len = CalcReferenceLength(sid, i, length);
stats.push_back(reference_len);
//ADD stats to totStats
std::transform(stats.begin(), stats.end(), totStats.begin(),
totStats.begin(), std::plus<int>());
}
entry.set(totStats);
}
std::vector<std::string> BleuDocScorer::splitDoc(const std::string& text)
{
std::vector<std::string> res;
uint index = 0;
std::string::size_type end;
while ((end = text.find(" \\n ", index)) != std::string::npos) {
res.push_back(text.substr(index,end-index));
index = end + 4;
}
return res;
}
statscore_t BleuDocScorer::calculateScore(const vector<int>& comps) const
{
UTIL_THROW_IF(comps.size() != kBleuNgramOrder * 2 + 1, util::Exception, "Error");
float logbleu = 0.0;
for (size_t 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 BleuDocScorer::CalcReferenceLength(size_t doc_id, size_t sentence_id, size_t length)
{
switch (m_ref_length_type) {
case AVERAGE:
return m_references[doc_id]->get().at(sentence_id)->CalcAverage();
break;
case CLOSEST:
return m_references[doc_id]->get().at(sentence_id)->CalcClosest(length);
break;
case SHORTEST:
return m_references[doc_id]->get().at(sentence_id)->CalcShortest();
break;
default:
cerr << "unknown reference types." << endl;
exit(1);
}
}
}
|
