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
|
#include "score.h"
namespace dtrain
{
Ngrams
make_ngrams(vector<WordID>& s, size_t N)
{
Ngrams ngrams;
vector<WordID> ng;
for (size_t i = 0; i < s.size(); i++) {
ng.clear();
for (size_t j = i; j < min(i+N, s.size()); j++) {
ng.push_back(s[j]);
ngrams[ng]++;
}
}
return ngrams;
}
NgramCounts
make_ngram_counts(vector<WordID> hyp, vector<WordID> ref, size_t N)
{
Ngrams hyp_ngrams = make_ngrams(hyp, N);
Ngrams ref_ngrams = make_ngrams(ref, N);
NgramCounts counts(N);
Ngrams::iterator it;
Ngrams::iterator ti;
for (it = hyp_ngrams.begin(); it != hyp_ngrams.end(); it++) {
ti = ref_ngrams.find(it->first);
if (ti != ref_ngrams.end()) {
counts.add(it->second, ti->second, it->first.size() - 1);
} else {
counts.add(it->second, 0, it->first.size() - 1);
}
}
return counts;
}
/*
* bleu
*
* as in "BLEU: a Method for Automatic Evaluation
* of Machine Translation"
* (Papineni et al. '02)
*
* NOTE: 0 if one n in {1..N} has 0 count
*/
double
brevity_penaly(const size_t hyp_len, const size_t ref_len)
{
if (hyp_len > ref_len) return 1;
return exp(1 - (double)ref_len/(double)hyp_len);
}
double
bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
size_t N, vector<float> weights )
{
if (hyp_len == 0 || ref_len == 0) return 0;
if (ref_len < N) N = ref_len;
float N_ = (float)N;
if (weights.empty())
{
for (size_t i = 0; i < N; i++) weights.push_back(1/N_);
}
double sum = 0;
for (size_t i = 0; i < N; i++) {
if (counts.clipped[i] == 0 || counts.sum[i] == 0) return 0;
sum += weights[i] * log((double)counts.clipped[i] / (double)counts.sum[i]);
}
return brevity_penaly(hyp_len, ref_len) * exp(sum);
}
/*
* 'stupid' bleu
*
* as in "ORANGE: a Method for Evaluating
* Automatic Evaluation Metrics
* for Machine Translation"
* (Lin & Och '04)
*
* NOTE: 0 iff no 1gram match
*/
double
stupid_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
size_t N, vector<float> weights )
{
if (hyp_len == 0 || ref_len == 0) return 0;
if (ref_len < N) N = ref_len;
float N_ = (float)N;
if (weights.empty())
{
for (size_t i = 0; i < N; i++) weights.push_back(1/N_);
}
double sum = 0;
float add = 0;
for (size_t i = 0; i < N; i++) {
if (i == 1) add = 1;
sum += weights[i] * log(((double)counts.clipped[i] + add) / ((double)counts.sum[i] + add));
}
return brevity_penaly(hyp_len, ref_len) * exp(sum);
}
/*
* smooth bleu
*
* as in "An End-to-End Discriminative Approach
* to Machine Translation"
* (Liang et al. '06)
*
* NOTE: max is 0.9375
*/
double
smooth_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
const size_t N, vector<float> weights )
{
if (hyp_len == 0 || ref_len == 0) return 0;
float N_ = (float)N;
if (weights.empty())
{
for (size_t i = 0; i < N; i++) weights.push_back(1/N_);
}
double sum = 0;
float j = 1;
for (size_t i = 0; i < N; i++) {
if (counts.clipped[i] == 0 || counts.sum[i] == 0) continue;
sum += exp((weights[i] * log((double)counts.clipped[i]/(double)counts.sum[i]))) / pow(2, N_-j+1);
j++;
}
return brevity_penaly(hyp_len, ref_len) * sum;
}
/*
* approx. bleu
*
* as in "Online Large-Margin Training of Syntactic
* and Structural Translation Features"
* (Chiang et al. '08)
*/
double
approx_bleu(NgramCounts& counts, const size_t hyp_len, const size_t ref_len,
const size_t N, vector<float> weights)
{
return brevity_penaly(hyp_len, ref_len)
* 0.9 * bleu(counts, hyp_len, ref_len, N, weights);
}
} // namespace
|
