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
|
#include "ff_csplit.h"
#include <set>
#include <cstring>
#include "Vocab.h"
#include "Ngram.h"
#include "sentence_metadata.h"
#include "lattice.h"
#include "tdict.h"
#include "freqdict.h"
#include "filelib.h"
#include "stringlib.h"
#include "tdict.h"
using namespace std;
struct BasicCSplitFeaturesImpl {
BasicCSplitFeaturesImpl(const string& param) :
word_count_(FD::Convert("WordCount")),
letters_sq_(FD::Convert("LettersSq")),
letters_sqrt_(FD::Convert("LettersSqrt")),
in_dict_(FD::Convert("InDict")),
in_dict_sub_word_(FD::Convert("InDictSubWord")),
short_(FD::Convert("Short")),
long_(FD::Convert("Long")),
oov_(FD::Convert("OOV")),
oov_sub_word_(FD::Convert("OOVSubWord")),
short_range_(FD::Convert("ShortRange")),
high_freq_(FD::Convert("HighFreq")),
med_freq_(FD::Convert("MedFreq")),
freq_(FD::Convert("Freq")),
fl1_(FD::Convert("FreqLen1")),
fl2_(FD::Convert("FreqLen2")),
bad_(FD::Convert("Bad")) {
vector<string> argv;
int argc = SplitOnWhitespace(param, &argv);
if (argc != 1 && argc != 2) {
cerr << "Expected: freqdict.txt [badwords.txt]\n";
abort();
}
freq_dict_.Load(argv[0]);
if (argc == 2) {
ReadFile rf(argv[1]);
istream& in = *rf.stream();
while(in) {
string badword;
in >> badword;
if (badword.empty()) continue;
bad_words_.insert(TD::Convert(badword));
}
}
}
void TraversalFeaturesImpl(const Hypergraph::Edge& edge,
const int src_word_size,
SparseVector<double>* features) const;
const int word_count_;
const int letters_sq_;
const int letters_sqrt_;
const int in_dict_;
const int in_dict_sub_word_;
const int short_;
const int long_;
const int oov_;
const int oov_sub_word_;
const int short_range_;
const int high_freq_;
const int med_freq_;
const int freq_;
const int fl1_;
const int fl2_;
const int bad_;
FreqDict freq_dict_;
set<WordID> bad_words_;
};
BasicCSplitFeatures::BasicCSplitFeatures(const string& param) :
pimpl_(new BasicCSplitFeaturesImpl(param)) {}
void BasicCSplitFeaturesImpl::TraversalFeaturesImpl(
const Hypergraph::Edge& edge,
const int src_word_length,
SparseVector<double>* features) const {
const bool subword = (edge.i_ > 0) || (edge.j_ < src_word_length);
features->set_value(word_count_, 1.0);
features->set_value(letters_sq_, (edge.j_ - edge.i_) * (edge.j_ - edge.i_));
features->set_value(letters_sqrt_, sqrt(edge.j_ - edge.i_));
const WordID word = edge.rule_->e_[1];
const char* sword = TD::Convert(word);
const int len = strlen(sword);
int cur = 0;
int chars = 0;
while(cur < len) {
cur += UTF8Len(sword[cur]);
++chars;
}
// these are corrections that attempt to make chars
// more like a phoneme count than a letter count, they
// are only really meaningful for german and should
// probably be gotten rid of
bool has_sch = strstr(sword, "sch");
bool has_ch = (!has_sch && strstr(sword, "ch"));
bool has_ie = strstr(sword, "ie");
bool has_zw = strstr(sword, "zw");
if (has_sch) chars -= 2;
if (has_ch) --chars;
if (has_ie) --chars;
if (has_zw) --chars;
float freq = freq_dict_.LookUp(word);
if (freq) {
features->set_value(freq_, freq);
features->set_value(in_dict_, 1.0);
if (subword) features->set_value(in_dict_sub_word_, 1.0);
} else {
features->set_value(oov_, 1.0);
if (subword) features->set_value(oov_sub_word_, 1.0);
freq = 99.0f;
}
if (bad_words_.count(word) != 0)
features->set_value(bad_, 1.0);
if (chars < 5)
features->set_value(short_, 1.0);
if (chars > 10)
features->set_value(long_, 1.0);
if (freq < 7.0f)
features->set_value(high_freq_, 1.0);
if (freq > 8.0f && freq < 10.f)
features->set_value(med_freq_, 1.0);
if (freq < 10.0f && chars < 5)
features->set_value(short_range_, 1.0);
// i don't understand these features, but they really help!
features->set_value(fl1_, sqrt(chars * freq));
features->set_value(fl2_, freq / chars);
}
void BasicCSplitFeatures::TraversalFeaturesImpl(
const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const std::vector<const void*>& ant_contexts,
SparseVector<double>* features,
SparseVector<double>* estimated_features,
void* out_context) const {
(void) smeta;
(void) ant_contexts;
(void) out_context;
(void) estimated_features;
if (edge.Arity() == 0) return;
if (edge.rule_->EWords() != 1) return;
pimpl_->TraversalFeaturesImpl(edge, smeta.GetSourceLattice().size(), features);
}
struct ReverseCharLMCSplitFeatureImpl {
ReverseCharLMCSplitFeatureImpl(const string& param) :
order_(5),
vocab_(TD::dict_),
ngram_(vocab_, order_) {
kBOS = vocab_.getIndex("<s>");
kEOS = vocab_.getIndex("</s>");
File file(param.c_str(), "r", 0);
assert(file);
cerr << "Reading " << order_ << "-gram LM from " << param << endl;
ngram_.read(file);
}
double LeftPhonotacticProb(const Lattice& inword, const int start) {
const int end = inword.size();
for (int i = 0; i < order_; ++i)
sc[i] = kBOS;
int sp = min(end - start, order_ - 1);
// cerr << "[" << start << "," << sp << "]\n";
int ci = (order_ - sp - 1);
int wi = start;
while (sp > 0) {
sc[ci] = inword[wi][0].label;
// cerr << " CHAR: " << TD::Convert(sc[ci]) << " ci=" << ci << endl;
++wi;
++ci;
--sp;
}
// cerr << " END ci=" << ci << endl;
sc[ci] = Vocab_None;
const double startprob = ngram_.wordProb(kEOS, sc);
// cerr << " PROB=" << startprob << endl;
return startprob;
}
private:
const int order_;
Vocab& vocab_;
VocabIndex kBOS;
VocabIndex kEOS;
Ngram ngram_;
VocabIndex sc[80];
};
ReverseCharLMCSplitFeature::ReverseCharLMCSplitFeature(const string& param) :
pimpl_(new ReverseCharLMCSplitFeatureImpl(param)),
fid_(FD::Convert("RevCharLM")) {}
void ReverseCharLMCSplitFeature::TraversalFeaturesImpl(
const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const std::vector<const void*>& ant_contexts,
SparseVector<double>* features,
SparseVector<double>* estimated_features,
void* out_context) const {
(void) ant_contexts;
(void) estimated_features;
(void) out_context;
if (edge.Arity() != 1) return;
if (edge.rule_->EWords() != 1) return;
const double lpp = pimpl_->LeftPhonotacticProb(smeta.GetSourceLattice(), edge.i_);
features->set_value(fid_, lpp);
#if 0
WordID neighbor_word = 0;
const WordID word = edge.rule_->e_[1];
const char* sword = TD::Convert(word);
const int len = strlen(sword);
int cur = 0;
int chars = 0;
while(cur < len) {
cur += UTF8Len(sword[cur]);
++chars;
}
if (chars > 4 && (sword[0] == 's' || sword[0] == 'n')) {
neighbor_word = TD::Convert(string(&sword[1]));
}
if (neighbor_word) {
float nfreq = freq_dict_.LookUp(neighbor_word);
cerr << "COMPARE: " << TD::Convert(word) << " & " << TD::Convert(neighbor_word) << endl;
if (!nfreq) nfreq = 99.0f;
features->set_value(fdoes_deletion_help_, (freq - nfreq));
}
#endif
}
|
