summaryrefslogtreecommitdiff
path: root/decoder/ff_spans.cc
blob: 17e1d650c36f8877ce2275bb721ee6a3d4beb074 (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
#include "ff_spans.h"

#include <sstream>
#include <cassert>
#include <cmath>

#include "filelib.h"
#include "stringlib.h"
#include "sentence_metadata.h"
#include "lattice.h"
#include "fdict.h"
#include "verbose.h"

using namespace std;

// log transform to make long spans cluster together
// but preserve differences
int SpanSizeTransform(unsigned span_size) {
  if (!span_size) return 0;
  return static_cast<int>(log(span_size+1) / log(1.39)) - 1;
}

SpanFeatures::SpanFeatures(const string& param) :
    kS(TD::Convert("S") * -1),
    kX(TD::Convert("X") * -1),
    use_collapsed_features_(false) {
  string mapfile = param;
  string valfile;
  vector<string> toks;
  Tokenize(param, ' ', &toks);
  if (toks.size() == 2) { mapfile = toks[0]; valfile = toks[1]; }
  if (mapfile.size() > 0) {
    int lc = 0;
    if (!SILENT) { cerr << "Reading word map for SpanFeatures from " << param << endl; }
    ReadFile rf(mapfile);
    istream& in = *rf.stream();
    string line;
    vector<WordID> v;
    while(in) {
      ++lc;
      getline(in, line);
      if (line.empty()) continue;
      v.clear();
      TD::ConvertSentence(line, &v);
      if (v.size() != 2) {
        cerr << "Error reading line " << lc << ": " << line << endl;
        abort();
      }
      word2class_[v[0]] = v[1];
    }
    word2class_[TD::Convert("BOS")] = TD::Convert("BOS");
    word2class_[TD::Convert("EOS")] = TD::Convert("EOS");
    oov_ = TD::Convert("OOV");
  }

  if (valfile.size() > 0) {
    use_collapsed_features_ = true;
    fid_beg_ = FD::Convert("SpanBegin");
    fid_end_ = FD::Convert("SpanEnd");
    fid_span_s_ = FD::Convert("SSpanContext");
    fid_span_ = FD::Convert("XSpanContext");
    ReadFile rf(valfile);
    if (!SILENT) { cerr << "  Loading span scores from " << valfile << endl; }
    istream& in = *rf.stream();
    string line;
    while(in) {
      getline(in, line);
      if (line.size() == 0 || line[0] == '#') { continue; }
      istringstream in(line);
      string feat_name;
      double weight;
      in >> feat_name >> weight;
      feat2val_[feat_name] = weight;
    }
  }
}

void SpanFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
                                         const Hypergraph::Edge& edge,
                                         const vector<const void*>& ant_contexts,
                                         SparseVector<double>* features,
                                         SparseVector<double>* estimated_features,
                                         void* context) const {
  assert(edge.j_ < end_span_ids_.size());
  assert(edge.j_ >= 0);
  assert(edge.i_ < beg_span_ids_.size());
  assert(edge.i_ >= 0);
  if (use_collapsed_features_) {
    features->set_value(fid_end_, end_span_vals_[edge.j_]);
    features->set_value(fid_beg_, beg_span_vals_[edge.i_]);
    if (edge.rule_->lhs_ == kS)
      features->set_value(fid_span_s_, span_vals_(edge.i_,edge.j_).second);
    else
      features->set_value(fid_span_, span_vals_(edge.i_,edge.j_).first);
  } else {  // non-collapsed features:
    features->set_value(end_span_ids_[edge.j_], 1);
    features->set_value(beg_span_ids_[edge.i_], 1);
    features->set_value(end_bigram_ids_[edge.j_], 1);
    features->set_value(beg_bigram_ids_[edge.i_], 1);
    if (edge.rule_->lhs_ == kS) {
      features->set_value(span_feats_(edge.i_,edge.j_).second, 1);
      features->set_value(len_span_feats_(edge.i_,edge.j_).second, 1);
    } else {
      features->set_value(span_feats_(edge.i_,edge.j_).first, 1);
      features->set_value(len_span_feats_(edge.i_,edge.j_).first, 1);
    }
  }
}

WordID SpanFeatures::MapIfNecessary(const WordID& w) const {
  if (word2class_.empty()) return w;
  map<WordID,WordID>::const_iterator it = word2class_.find(w);
  if (it == word2class_.end()) return oov_;
  return it->second;
}

void SpanFeatures::PrepareForInput(const SentenceMetadata& smeta) {
  const Lattice& lattice = smeta.GetSourceLattice();
  const WordID eos = TD::Convert("EOS");  // right of the last source word
  const WordID bos = TD::Convert("BOS");  // left of the first source word
  beg_span_ids_.resize(lattice.size() + 1);
  end_span_ids_.resize(lattice.size() + 1);
  span_feats_.resize(lattice.size() + 1, lattice.size() + 1);
  beg_bigram_ids_.resize(lattice.size() + 1);
  end_bigram_ids_.resize(lattice.size() + 1);
  len_span_feats_.resize(lattice.size() + 1, lattice.size() + 1);
  if (use_collapsed_features_) {
    beg_span_vals_.resize(lattice.size() + 1);
    end_span_vals_.resize(lattice.size() + 1);
    span_vals_.resize(lattice.size() + 1, lattice.size() + 1);
  }
  for (int i = 0; i <= lattice.size(); ++i) {
    WordID word = eos;
    WordID bword = bos;
    if (i > 0)
      bword = lattice[i-1][0].label;
    bword = MapIfNecessary(bword);
    if (i < lattice.size())
      word = lattice[i][0].label;  // rather arbitrary for lattices
    word = MapIfNecessary(word);
    ostringstream sfid;
    sfid << "ES:" << TD::Convert(word);
    end_span_ids_[i] = FD::Convert(sfid.str());
    ostringstream esbiid;
    esbiid << "EBI:" << TD::Convert(bword) << "_" << TD::Convert(word);
    end_bigram_ids_[i] = FD::Convert(esbiid.str());
    ostringstream bsbiid;
    bsbiid << "BBI:" << TD::Convert(bword) << "_" << TD::Convert(word);
    beg_bigram_ids_[i] = FD::Convert(bsbiid.str());
    ostringstream bfid;
    bfid << "BS:" << TD::Convert(bword);
    beg_span_ids_[i] = FD::Convert(bfid.str());
    if (use_collapsed_features_) {
      end_span_vals_[i] = feat2val_[sfid.str()];
      beg_span_vals_[i] = feat2val_[bfid.str()];
    }
  }
  for (int i = 0; i <= lattice.size(); ++i) {
    WordID bword = bos;
    if (i > 0)
      bword = lattice[i-1][0].label;
    bword = MapIfNecessary(bword);
    for (int j = 0; j <= lattice.size(); ++j) {
      WordID word = eos;
      if (j < lattice.size())
        word = lattice[j][0].label;
      word = MapIfNecessary(word);
      ostringstream pf;
      pf << "S:" << TD::Convert(bword) << "_" << TD::Convert(word);
      span_feats_(i,j).first = FD::Convert(pf.str());
      span_feats_(i,j).second = FD::Convert("S_" + pf.str());
      ostringstream lf;
      const unsigned span_size = (i < j ? j - i : i - j);
      lf << "LS:" << SpanSizeTransform(span_size) << "_" << TD::Convert(bword) << "_" << TD::Convert(word);
      len_span_feats_(i,j).first = FD::Convert(lf.str());
      len_span_feats_(i,j).second = FD::Convert("S_" + lf.str());
      if (use_collapsed_features_) {
        span_vals_(i,j).first = feat2val_[pf.str()];
        span_vals_(i,j).second = feat2val_["S_" + pf.str()];
      }
    }
  } 
}

inline bool IsArity2RuleReordered(const TRule& rule) {
  const vector<WordID>& e = rule.e_;
  for (int i = 0; i < e.size(); ++i) {
    if (e[i] <= 0) { return e[i] < 0; }
  }
  cerr << "IsArity2RuleReordered failed on:\n" << rule.AsString() << endl;
  abort();
}

// Chiang, Marton, Resnik 2008 "fine-grained" reordering features
CMR2008ReorderingFeatures::CMR2008ReorderingFeatures(const std::string& param) :
    kS(TD::Convert("S") * -1),
    use_collapsed_features_(false) {
  if (param.size() > 0) {
    use_collapsed_features_ = true;
    assert(!"not implemented"); // TODO
  } else {
    unconditioned_fids_.first = FD::Convert("CMRMono");
    unconditioned_fids_.second = FD::Convert("CMRReorder");
    fids_.resize(16); fids_[0].first = fids_[0].second = -1;
    // since I use a log transform, I go a bit higher than David, who bins everything > 10
    for (int span_size = 1; span_size <= 15; ++span_size) {
      ostringstream m, r;
      m << "CMRMono_" << SpanSizeTransform(span_size);
      fids_[span_size].first = FD::Convert(m.str());
      r << "CMRReorder_" << SpanSizeTransform(span_size);
      fids_[span_size].second = FD::Convert(r.str());
    }
  }
}

void CMR2008ReorderingFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
                                         const Hypergraph::Edge& edge,
                                         const vector<const void*>& ant_contexts,
                                         SparseVector<double>* features,
                                         SparseVector<double>* estimated_features,
                                         void* context) const {
  if (edge.Arity() != 2) return;
  if (edge.rule_->lhs_ == kS) return;
  assert(edge.i_ >= 0);
  assert(edge.j_ > edge.i_);
  const bool is_reordered = IsArity2RuleReordered(*edge.rule_);
  const unsigned span_size = edge.j_ - edge.i_;
  if (use_collapsed_features_) {
    assert(!"not impl"); // TODO
  } else {
    if (is_reordered) {
      features->set_value(unconditioned_fids_.second, 1.0);
      features->set_value(fids_[span_size].second, 1.0);
    } else {
      features->set_value(unconditioned_fids_.first, 1.0);
      features->set_value(fids_[span_size].first, 1.0);
    }
  }
}