summaryrefslogtreecommitdiff
path: root/training/dtrain/score_net_interface.h
blob: 58357cf607575f4a053477e07726101dc8a482b3 (plain)
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
#ifndef _DTRAIN_SCORE_NET_INTERFACE_H_
#define _DTRAIN_SCORE_NET_INTERFACE_H_

#include "dtrain.h"

namespace dtrain
{

struct NgramCounts
{
  size_t N_;
  map<size_t, weight_t> clipped_;
  map<size_t, weight_t> sum_;

  NgramCounts(const size_t N) : N_(N) { Zero(); }

  inline void
  operator+=(const NgramCounts& rhs)
  {
    if (rhs.N_ > N_) Resize(rhs.N_);
    for (size_t i = 0; i < N_; i++) {
      this->clipped_[i] += rhs.clipped_.find(i)->second;
      this->sum_[i] += rhs.sum_.find(i)->second;
    }
  }

  inline const NgramCounts
  operator+(const NgramCounts &other) const
  {
    NgramCounts result = *this;
    result += other;

    return result;
  }

  inline void
  Add(const size_t count, const size_t ref_count, const size_t i)
  {
    assert(i < N_);
    if (count > ref_count) {
      clipped_[i] += ref_count;
    } else {
      clipped_[i] += count;
    }
    sum_[i] += count;
  }

  inline void
  Zero()
  {
    for (size_t i = 0; i < N_; i++) {
      clipped_[i] = 0.;
      sum_[i] = 0.;
    }
  }

  inline void
  Resize(size_t N)
  {
    if (N == N_) return;
    else if (N > N_) {
      for (size_t i = N_; i < N; i++) {
        clipped_[i] = 0.;
        sum_[i] = 0.;
      }
    } else { // N < N_
      for (size_t i = N_-1; i > N-1; i--) {
        clipped_.erase(i);
        sum_.erase(i);
      }
    }
    N_ = N;
  }
};

typedef map<vector<WordID>, size_t> Ngrams;

inline Ngrams
MakeNgrams(const vector<WordID>& s, const 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;
}

inline NgramCounts
MakeNgramCounts(const vector<WordID>& hyp,
                const vector<Ngrams>& ref,
                const size_t N)
{
  Ngrams hyp_ngrams = MakeNgrams(hyp, N);
  NgramCounts counts(N);
  Ngrams::iterator it, ti;
  for (it = hyp_ngrams.begin(); it != hyp_ngrams.end(); it++) {
    size_t max_ref_count = 0;
    for (auto r: ref) {
      ti = r.find(it->first);
      if (ti != r.end())
        max_ref_count = max(max_ref_count, ti->second);
    }
    counts.Add(it->second, min(it->second, max_ref_count), it->first.size()-1);
  }

  return counts;
}

/*
 * per-sentence BLEU
 * as in "Optimizing for Sentence-Level BLEU+1
 *        Yields Short Translations"
 * (Nakov et al. '12)
 *
 * [simply add 1 to reference length for calculation of BP]
 *
 */
struct PerSentenceBleuScorer
{
  const size_t     N_;
  vector<weight_t> w_;

  PerSentenceBleuScorer(size_t n) : N_(n)
  {
    for (size_t i = 1; i <= N_; i++)
      w_.push_back(1.0/N_);
  }

  inline weight_t
  BrevityPenalty(const size_t hl, const size_t rl)
  {
    if (hl > rl)
      return 1;

    return exp(1 - (weight_t)rl/hl);
  }

  inline size_t
  BestMatchLength(const size_t hl,
                  const vector<size_t>& ref_ls)
  {
    size_t m;
    if (ref_ls.size() == 1)  {
      m = ref_ls.front();
    } else {
      size_t i = 0, best_idx = 0;
      size_t best = numeric_limits<size_t>::max();
      for (auto l: ref_ls) {
        size_t d = abs(hl-l);
        if (d < best) {
          best_idx = i;
          best = d;
        }
        i += 1;
      }
      m = ref_ls[best_idx];
    }

    return m;
  }

  weight_t
  Score(const vector<WordID>& hyp,
        const vector<Ngrams>& ref_ngs,
        const vector<size_t>& ref_ls)
  {
    size_t hl = hyp.size(), rl = 0;
    if (hl == 0) return 0.;
    rl = BestMatchLength(hl, ref_ls);
    if (rl == 0) return 0.;
    NgramCounts counts = MakeNgramCounts(hyp, ref_ngs, N_);
    size_t M = N_;
    vector<weight_t> v = w_;
    if (rl < N_) {
      M = rl;
      for (size_t i = 0; i < M; i++) v[i] = 1/((weight_t)M);
    }
    weight_t sum = 0, add = 0;
    for (size_t i = 0; i < M; i++) {
      if (i == 0 && (counts.sum_[i] == 0 || counts.clipped_[i] == 0)) return 0.;
      if (i > 0) add = 1;
      sum += v[i] * log(((weight_t)counts.clipped_[i] + add)
                        / ((counts.sum_[i] + add)));
    }

    //return  BrevityPenalty(hl, rl+1) * exp(sum);
    return  BrevityPenalty(hl, rl) * exp(sum);
  }
};

} // namespace

#endif