summaryrefslogtreecommitdiff
path: root/mteval/ns.cc
blob: 6139757d9f27a5ca17039907f5c14a834b136063 (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
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
#include "ns.h"
#include "ns_ter.h"
#include "ns_ext.h"
#include "ns_comb.h"

#include <cassert>
#include <cmath>
#include <cstdlib>
#include <iostream>
#include <sstream>

#include "tdict.h"
#include "stringlib.h"

using namespace std;
using boost::shared_ptr;

map<string, EvaluationMetric*> EvaluationMetric::instances_;

SegmentEvaluator::~SegmentEvaluator() {}
EvaluationMetric::~EvaluationMetric() {}

struct DefaultSegmentEvaluator : public SegmentEvaluator {
  DefaultSegmentEvaluator(const vector<vector<WordID> >& refs, const EvaluationMetric* em) : refs_(refs), em_(em) {}
  void Evaluate(const vector<WordID>& hyp, SufficientStats* out) const {
    em_->ComputeSufficientStatistics(hyp, refs_, out);
    out->id_ = em_->MetricId();
  }
  const vector<vector<WordID> > refs_;
  const EvaluationMetric* em_;
};

shared_ptr<SegmentEvaluator> EvaluationMetric::CreateSegmentEvaluator(const vector<vector<WordID> >& refs) const {
  return shared_ptr<SegmentEvaluator>(new DefaultSegmentEvaluator(refs, this));
}

#define MAX_SS_VECTOR_SIZE 50
unsigned EvaluationMetric::SufficientStatisticsVectorSize() const {
  return MAX_SS_VECTOR_SIZE;
}

void EvaluationMetric::ComputeSufficientStatistics(const vector<WordID>&,
                                                   const vector<vector<WordID> >&,
                                                   SufficientStats*) const {
  cerr << "Base class ComputeSufficientStatistics should not be called.\n";
  abort();
}

string EvaluationMetric::DetailedScore(const SufficientStats& stats) const {
  ostringstream os;
  os << MetricId() << "=" << ComputeScore(stats);
  return os.str();
}

enum BleuType { IBM, Koehn, NIST };
template <unsigned int N = 4u, BleuType BrevityType = IBM>
struct BleuSegmentEvaluator : public SegmentEvaluator {
  BleuSegmentEvaluator(const vector<vector<WordID> >& refs, const EvaluationMetric* em) : evaluation_metric(em) {
    assert(refs.size() > 0);
    float tot = 0;
    int smallest = 9999999;
    for (vector<vector<WordID> >::const_iterator ci = refs.begin();
         ci != refs.end(); ++ci) {
      lengths_.push_back(ci->size());
      tot += lengths_.back();
      if (lengths_.back() < smallest) smallest = lengths_.back();
      CountRef(*ci);
    }
    if (BrevityType == Koehn)
      lengths_[0] = tot / refs.size();
    if (BrevityType == NIST)
      lengths_[0] = smallest;
  }

  void Evaluate(const vector<WordID>& hyp, SufficientStats* out) const {
    out->fields.resize(N + N + 2);
    out->id_ = evaluation_metric->MetricId();
    for (unsigned i = 0; i < N+N+2; ++i) out->fields[i] = 0;

    ComputeNgramStats(hyp, &out->fields[0], &out->fields[N], true);
    float& hyp_len = out->fields[2*N];
    float& ref_len = out->fields[2*N + 1];
    hyp_len = hyp.size();
    ref_len = lengths_[0];
    if (lengths_.size() > 1 && BrevityType == IBM) {
      float bestd = 2000000;
      float hl = hyp.size();
      float bl = -1;
      for (vector<float>::const_iterator ci = lengths_.begin(); ci != lengths_.end(); ++ci) {
        if (fabs(*ci - hl) < bestd) {
          bestd = fabs(*ci - hl);
          bl = *ci;
        }
      }
      ref_len = bl;
    }
  }

  struct NGramCompare {
    int operator() (const vector<WordID>& a, const vector<WordID>& b) {
      const size_t as = a.size();
      const size_t bs = b.size();
      const size_t s = (as < bs ? as : bs);
      for (size_t i = 0; i < s; ++i) {
         int d = a[i] - b[i];
         if (d < 0) return true;
         if (d > 0) return false;
      }
      return as < bs;
    }
  };
  typedef map<vector<WordID>, pair<int,int>, NGramCompare> NGramCountMap;

  void CountRef(const vector<WordID>& ref) {
    NGramCountMap tc;
    vector<WordID> ngram(N);
    int s = ref.size();
    for (int j=0; j<s; ++j) {
      int remaining = s-j;
      int k = (N < remaining ? N : remaining);
      ngram.clear();
      for (int i=1; i<=k; ++i) {
        ngram.push_back(ref[j + i - 1]);
        tc[ngram].first++;
      }
    }
    for (typename NGramCountMap::iterator i = tc.begin(); i != tc.end(); ++i) {
      pair<int,int>& p = ngrams_[i->first];
      if (p.first < i->second.first)
        p = i->second;
    }
  }

  void ComputeNgramStats(const vector<WordID>& sent,
                         float* correct,  // N elements reserved
                         float* hyp,      // N elements reserved
                         bool clip_counts = true) const {
    vector<WordID> ngram(N);
    *correct *= 0;
    *hyp *= 0;
    int s = sent.size();
    for (int j=0; j<s; ++j) {
      int remaining = s-j;
      int k = (N < remaining ? N : remaining);
      ngram.clear();
      for (int i=1; i<=k; ++i) {
        ngram.push_back(sent[j + i - 1]);
        pair<int,int>& p = ngrams_[ngram];
        if(clip_counts){
          if (p.second < p.first) {
            ++p.second;
            correct[i-1]++;
          }
        } else {
          ++p.second;
          correct[i-1]++;
        }
        // if the 1 gram isn't found, don't try to match don't need to match any 2- 3- .. grams:
        if (!p.first) {
          for (; i<=k; ++i)
            hyp[i-1]++;
        } else {
          hyp[i-1]++;
        }
      }
    }
  }

  const EvaluationMetric* evaluation_metric;
  vector<float> lengths_;
  mutable NGramCountMap ngrams_;
};

template <unsigned int N = 4u, BleuType BrevityType = IBM>
struct BleuMetric : public EvaluationMetric {
  BleuMetric() : EvaluationMetric("IBM_BLEU") {}
  unsigned SufficientStatisticsVectorSize() const { return N*2 + 2; }
  shared_ptr<SegmentEvaluator> CreateSegmentEvaluator(const vector<vector<WordID> >& refs) const {
    return shared_ptr<SegmentEvaluator>(new BleuSegmentEvaluator<N,BrevityType>(refs, this));
  }
  float ComputeBreakdown(const SufficientStats& stats, float* bp, vector<float>* out) const {
    if (out) { out->clear(); }
    float log_bleu = 0;
    int count = 0;
    for (int i = 0; i < N; ++i) {
      if (stats.fields[i+N] > 0) {
        float cor_count = stats.fields[i];  // correct_ngram_hit_counts[i];
        // smooth bleu
        if (!cor_count) { cor_count = 0.01; }
        float lprec = log(cor_count) - log(stats.fields[i+N]); // log(hyp_ngram_counts[i]);
        if (out) out->push_back(exp(lprec));
        log_bleu += lprec;
        ++count;
      }
    }
    log_bleu /= count;
    float lbp = 0.0;
    const float& hyp_len = stats.fields[2*N];
    const float& ref_len = stats.fields[2*N + 1];
    if (hyp_len < ref_len)
      lbp = (hyp_len - ref_len) / hyp_len;
    log_bleu += lbp;
    if (bp) *bp = exp(lbp);
    return exp(log_bleu);
  }
  string DetailedScore(const SufficientStats& stats) const {
    char buf[2000];
    vector<float> precs(N);
    float bp;
    float bleu = ComputeBreakdown(stats, &bp, &precs);
    sprintf(buf, "BLEU = %.2f, %.1f|%.1f|%.1f|%.1f (brev=%.3f)",
       bleu*100.0,
       precs[0]*100.0,
       precs[1]*100.0,
       precs[2]*100.0,
       precs[3]*100.0,
       bp);
    return buf;
  }
  float ComputeScore(const SufficientStats& stats) const {
    return ComputeBreakdown(stats, NULL, NULL);
  }
};

EvaluationMetric* EvaluationMetric::Instance(const string& imetric_id) {
  static bool is_first = true;
  if (is_first) {
    instances_["NULL"] = NULL;
    is_first = false;
  }
  const string metric_id = UppercaseString(imetric_id);

  map<string, EvaluationMetric*>::iterator it = instances_.find(metric_id);
  if (it == instances_.end()) {
    EvaluationMetric* m = NULL; 
    if        (metric_id == "IBM_BLEU") {
      m = new BleuMetric<4, IBM>;
    } else if (metric_id == "NIST_BLEU") {
      m = new BleuMetric<4, NIST>;
    } else if (metric_id == "KOEHN_BLEU") {
      m = new BleuMetric<4, Koehn>;
    } else if (metric_id == "TER") {
      m = new TERMetric;
    } else if (metric_id == "METEOR") {
      m = new ExternalMetric("METEOR", "java -Xmx1536m -jar /Users/cdyer/software/meteor/meteor-1.3.jar - - -mira -lower -t tune -l en");
    } else if (metric_id.find("COMB:") == 0) {
      m = new CombinationMetric(metric_id);
    } else {
      cerr << "Implement please: " << metric_id << endl;
      abort();
    }
    if (m->MetricId() != metric_id) {
      cerr << "Registry error: " << metric_id << " vs. " << m->MetricId() << endl;
      abort();
    }
    return instances_[metric_id] = m;
  } else {
    return it->second;
  }
}

SufficientStats::SufficientStats(const string& encoded) {
  istringstream is(encoded);
  is >> id_;
  float val;
  while(is >> val)
    fields.push_back(val);
}

void SufficientStats::Encode(string* out) const {
  ostringstream os;
  if (id_.size() > 0)
    os << id_;
  else
    os << "NULL";
  for (unsigned i = 0; i < fields.size(); ++i)
    os << ' ' << fields[i];
  *out = os.str();
}