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
|
#include "ff_source_syntax.h"
#include <sstream>
#include <stack>
#ifdef HAVE_CXX11
# include <unordered_set>
#else
# include <tr1/unordered_set>
namespace std { using std::tr1::unordered_set; }
#endif
#include "hg.h"
#include "sentence_metadata.h"
#include "array2d.h"
#include "filelib.h"
using namespace std;
// implements the source side syntax features described in Blunsom et al. (EMNLP 2008)
// source trees must be represented in Penn Treebank format, e.g.
// (S (NP John) (VP (V left)))
// log transform to make long spans cluster together
// but preserve differences
inline int SpanSizeTransform(unsigned span_size) {
if (!span_size) return 0;
return static_cast<int>(log(span_size+1) / log(1.39)) - 1;
}
struct SourceSyntaxFeaturesImpl {
SourceSyntaxFeaturesImpl() {}
void InitializeGrids(const string& tree, unsigned src_len) {
assert(tree.size() > 0);
//fids_cat.clear();
fids_ef.clear();
src_tree.clear();
//fids_cat.resize(src_len, src_len + 1);
fids_ef.resize(src_len, src_len + 1);
src_tree.resize(src_len, src_len + 1, TD::Convert("XX"));
ParseTreeString(tree, src_len);
}
void ParseTreeString(const string& tree, unsigned src_len) {
stack<pair<int, WordID> > stk; // first = i, second = category
pair<int, WordID> cur_cat; cur_cat.first = -1;
unsigned i = 0;
unsigned p = 0;
while(p < tree.size()) {
const char cur = tree[p];
if (cur == '(') {
stk.push(cur_cat);
++p;
unsigned k = p + 1;
while (k < tree.size() && tree[k] != ' ') { ++k; }
cur_cat.first = i;
cur_cat.second = TD::Convert(tree.substr(p, k - p));
// cerr << "NT: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
p = k + 1;
} else if (cur == ')') {
unsigned k = p;
while (k < tree.size() && tree[k] == ')') { ++k; }
const unsigned num_closes = k - p;
for (unsigned ci = 0; ci < num_closes; ++ci) {
// cur_cat.second spans from cur_cat.first to i
// cerr << TD::Convert(cur_cat.second) << " from " << cur_cat.first << " to " << i << endl;
// NOTE: unary rule chains end up being labeled with the top-most category
src_tree(cur_cat.first, i) = cur_cat.second;
cur_cat = stk.top();
stk.pop();
}
p = k;
while (p < tree.size() && (tree[p] == ' ' || tree[p] == '\t')) { ++p; }
} else if (cur == ' ' || cur == '\t') {
cerr << "Unexpected whitespace in: " << tree << endl;
abort();
} else { // terminal symbol
unsigned k = p + 1;
do {
while (k < tree.size() && tree[k] != ')' && tree[k] != ' ') { ++k; }
// cerr << "TERM: '" << tree.substr(p, k-p) << "' (i=" << i << ")\n";
++i;
assert(i <= src_len);
while (k < tree.size() && tree[k] == ' ') { ++k; }
p = k;
} while (p < tree.size() && tree[p] != ')');
}
}
// cerr << "i=" << i << " src_len=" << src_len << endl;
assert(i == src_len); // make sure tree specified in src_tree is
// the same length as the source sentence
}
WordID FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector<double>* feats) {
//cerr << "fire features: " << rule.AsString() << " for " << i << "," << j << endl;
const WordID lhs = src_tree(i,j);
//int& fid_cat = fids_cat(i,j);
int& fid_ef = fids_ef(i,j)[&rule];
if (fid_ef <= 0) {
ostringstream os;
//ostringstream os2;
os << "SYN:" << TD::Convert(lhs);
//os2 << "SYN:" << TD::Convert(lhs) << '_' << SpanSizeTransform(j - i);
//fid_cat = FD::Convert(os2.str());
os << ':';
unsigned ntc = 0;
for (unsigned k = 0; k < rule.f_.size(); ++k) {
if (k > 0) os << '_';
int fj = rule.f_[k];
if (fj <= 0) {
os << '[' << TD::Convert(ants[ntc++]) << ']';
} else {
os << TD::Convert(fj);
}
}
os << ':';
for (unsigned k = 0; k < rule.e_.size(); ++k) {
const int ei = rule.e_[k];
if (k > 0) os << '_';
if (ei <= 0)
os << '[' << (1-ei) << ']';
else
os << TD::Convert(ei);
}
fid_ef = FD::Convert(os.str());
}
//if (fid_cat > 0)
// feats->set_value(fid_cat, 1.0);
if (fid_ef > 0)
feats->set_value(fid_ef, 1.0);
return lhs;
}
Array2D<WordID> src_tree; // src_tree(i,j) NT = type
// mutable Array2D<int> fids_cat; // this tends to overfit baddly
mutable Array2D<map<const TRule*, int> > fids_ef; // fires for fully lexicalized
};
SourceSyntaxFeatures::SourceSyntaxFeatures(const string& param) :
FeatureFunction(sizeof(WordID)) {
impl = new SourceSyntaxFeaturesImpl;
}
SourceSyntaxFeatures::~SourceSyntaxFeatures() {
delete impl;
impl = NULL;
}
void SourceSyntaxFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const vector<const void*>& ant_contexts,
SparseVector<double>* features,
SparseVector<double>* estimated_features,
void* context) const {
WordID ants[8];
for (unsigned i = 0; i < ant_contexts.size(); ++i)
ants[i] = *static_cast<const WordID*>(ant_contexts[i]);
*static_cast<WordID*>(context) =
impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
}
void SourceSyntaxFeatures::PrepareForInput(const SentenceMetadata& smeta) {
impl->InitializeGrids(smeta.GetSGMLValue("src_tree"), smeta.GetSourceLength());
}
struct SourceSpanSizeFeaturesImpl {
SourceSpanSizeFeaturesImpl() {}
void InitializeGrids(unsigned src_len) {
fids.clear();
fids.resize(src_len, src_len + 1);
}
int FireFeatures(const TRule& rule, const int i, const int j, const WordID* ants, SparseVector<double>* feats) {
if (rule.Arity() > 0) {
int& fid = fids(i,j)[&rule];
if (fid <= 0) {
ostringstream os;
os << "SSS:";
unsigned ntc = 0;
for (unsigned k = 0; k < rule.f_.size(); ++k) {
if (k > 0) os << '_';
int fj = rule.f_[k];
if (fj <= 0) {
os << '[' << TD::Convert(-fj) << ants[ntc++] << ']';
} else {
os << TD::Convert(fj);
}
}
os << ':';
for (unsigned k = 0; k < rule.e_.size(); ++k) {
const int ei = rule.e_[k];
if (k > 0) os << '_';
if (ei <= 0)
os << '[' << (1-ei) << ']';
else
os << TD::Convert(ei);
}
fid = FD::Convert(os.str());
}
if (fid > 0)
feats->set_value(fid, 1.0);
}
return SpanSizeTransform(j - i);
}
mutable Array2D<map<const TRule*, int> > fids;
};
SourceSpanSizeFeatures::SourceSpanSizeFeatures(const string& param) :
FeatureFunction(sizeof(char)) {
impl = new SourceSpanSizeFeaturesImpl;
}
SourceSpanSizeFeatures::~SourceSpanSizeFeatures() {
delete impl;
impl = NULL;
}
void SourceSpanSizeFeatures::TraversalFeaturesImpl(const SentenceMetadata& smeta,
const Hypergraph::Edge& edge,
const vector<const void*>& ant_contexts,
SparseVector<double>* features,
SparseVector<double>* estimated_features,
void* context) const {
int ants[8];
for (unsigned i = 0; i < ant_contexts.size(); ++i)
ants[i] = *static_cast<const char*>(ant_contexts[i]);
*static_cast<char*>(context) =
impl->FireFeatures(*edge.rule_, edge.i_, edge.j_, ants, features);
}
void SourceSpanSizeFeatures::PrepareForInput(const SentenceMetadata& smeta) {
impl->InitializeGrids(smeta.GetSourceLength());
}
|
