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 "rule_factory.h"
#include <cassert>
#include <memory>
#include <queue>
#include <vector>
#include "grammar.h"
#include "intersector.h"
#include "matchings_finder.h"
#include "matching_comparator.h"
#include "phrase.h"
#include "rule.h"
#include "rule_extractor.h"
#include "sampler.h"
#include "scorer.h"
#include "suffix_array.h"
#include "vocabulary.h"
using namespace std;
using namespace tr1;
struct State {
State(int start, int end, const vector<int>& phrase,
const vector<int>& subpatterns_start, shared_ptr<TrieNode> node,
bool starts_with_x) :
start(start), end(end), phrase(phrase),
subpatterns_start(subpatterns_start), node(node),
starts_with_x(starts_with_x) {}
int start, end;
vector<int> phrase, subpatterns_start;
shared_ptr<TrieNode> node;
bool starts_with_x;
};
HieroCachingRuleFactory::HieroCachingRuleFactory(
shared_ptr<SuffixArray> source_suffix_array,
shared_ptr<DataArray> target_data_array,
shared_ptr<Alignment> alignment,
const shared_ptr<Vocabulary>& vocabulary,
shared_ptr<Precomputation> precomputation,
shared_ptr<Scorer> scorer,
int min_gap_size,
int max_rule_span,
int max_nonterminals,
int max_rule_symbols,
int max_samples,
bool use_baeza_yates,
bool require_tight_phrases) :
vocabulary(vocabulary),
scorer(scorer),
min_gap_size(min_gap_size),
max_rule_span(max_rule_span),
max_nonterminals(max_nonterminals),
max_chunks(max_nonterminals + 1),
max_rule_symbols(max_rule_symbols) {
matchings_finder = make_shared<MatchingsFinder>(source_suffix_array);
shared_ptr<MatchingComparator> comparator =
make_shared<MatchingComparator>(min_gap_size, max_rule_span);
intersector = make_shared<Intersector>(vocabulary, precomputation,
source_suffix_array, comparator, use_baeza_yates);
phrase_builder = make_shared<PhraseBuilder>(vocabulary);
rule_extractor = make_shared<RuleExtractor>(source_suffix_array->GetData(),
target_data_array, alignment, phrase_builder, scorer, vocabulary,
max_rule_span, min_gap_size, max_nonterminals, max_rule_symbols, true,
false, require_tight_phrases);
sampler = make_shared<Sampler>(source_suffix_array, max_samples);
}
Grammar HieroCachingRuleFactory::GetGrammar(const vector<int>& word_ids) {
// Clear cache for every new sentence.
trie.Reset();
shared_ptr<TrieNode> root = trie.GetRoot();
int first_x = vocabulary->GetNonterminalIndex(1);
shared_ptr<TrieNode> x_root(new TrieNode(root));
root->AddChild(first_x, x_root);
queue<State> states;
for (size_t i = 0; i < word_ids.size(); ++i) {
states.push(State(i, i, vector<int>(), vector<int>(1, i), root, false));
}
for (size_t i = min_gap_size; i < word_ids.size(); ++i) {
states.push(State(i - min_gap_size, i, vector<int>(1, first_x),
vector<int>(1, i), x_root, true));
}
vector<Rule> rules;
while (!states.empty()) {
State state = states.front();
states.pop();
shared_ptr<TrieNode> node = state.node;
vector<int> phrase = state.phrase;
int word_id = word_ids[state.end];
phrase.push_back(word_id);
Phrase next_phrase = phrase_builder->Build(phrase);
shared_ptr<TrieNode> next_node;
if (CannotHaveMatchings(node, word_id)) {
if (!node->HasChild(word_id)) {
node->AddChild(word_id, shared_ptr<TrieNode>());
}
continue;
}
if (RequiresLookup(node, word_id)) {
shared_ptr<TrieNode> next_suffix_link =
node->suffix_link->GetChild(word_id);
if (state.starts_with_x) {
// If the phrase starts with a non terminal, we simply use the matchings
// from the suffix link.
next_node = shared_ptr<TrieNode>(new TrieNode(
next_suffix_link, next_phrase, next_suffix_link->matchings));
} else {
PhraseLocation phrase_location;
if (next_phrase.Arity() > 0) {
phrase_location = intersector->Intersect(
node->phrase,
node->matchings,
next_suffix_link->phrase,
next_suffix_link->matchings,
next_phrase);
} else {
phrase_location = matchings_finder->Find(
node->matchings,
vocabulary->GetTerminalValue(word_id),
state.phrase.size());
}
if (phrase_location.IsEmpty()) {
continue;
}
next_node = shared_ptr<TrieNode>(new TrieNode(
next_suffix_link, next_phrase, phrase_location));
}
node->AddChild(word_id, next_node);
// Automatically adds a trailing non terminal if allowed. Simply copy the
// matchings from the prefix node.
AddTrailingNonterminal(phrase, next_phrase, next_node,
state.starts_with_x);
if (!state.starts_with_x) {
PhraseLocation sample = sampler->Sample(next_node->matchings);
vector<Rule> new_rules =
rule_extractor->ExtractRules(next_phrase, sample);
rules.insert(rules.end(), new_rules.begin(), new_rules.end());
}
} else {
next_node = node->GetChild(word_id);
}
vector<State> new_states = ExtendState(word_ids, state, phrase, next_phrase,
next_node);
for (State new_state: new_states) {
states.push(new_state);
}
}
return Grammar(rules, scorer->GetFeatureNames());
}
bool HieroCachingRuleFactory::CannotHaveMatchings(
shared_ptr<TrieNode> node, int word_id) {
if (node->HasChild(word_id) && node->GetChild(word_id) == NULL) {
return true;
}
shared_ptr<TrieNode> suffix_link = node->suffix_link;
return suffix_link != NULL && suffix_link->GetChild(word_id) == NULL;
}
bool HieroCachingRuleFactory::RequiresLookup(
shared_ptr<TrieNode> node, int word_id) {
return !node->HasChild(word_id);
}
void HieroCachingRuleFactory::AddTrailingNonterminal(
vector<int> symbols,
const Phrase& prefix,
const shared_ptr<TrieNode>& prefix_node,
bool starts_with_x) {
if (prefix.Arity() >= max_nonterminals) {
return;
}
int var_id = vocabulary->GetNonterminalIndex(prefix.Arity() + 1);
symbols.push_back(var_id);
Phrase var_phrase = phrase_builder->Build(symbols);
int suffix_var_id = vocabulary->GetNonterminalIndex(
prefix.Arity() + starts_with_x == 0);
shared_ptr<TrieNode> var_suffix_link =
prefix_node->suffix_link->GetChild(suffix_var_id);
prefix_node->AddChild(var_id, shared_ptr<TrieNode>(new TrieNode(
var_suffix_link, var_phrase, prefix_node->matchings)));
}
vector<State> HieroCachingRuleFactory::ExtendState(
const vector<int>& word_ids,
const State& state,
vector<int> symbols,
const Phrase& phrase,
const shared_ptr<TrieNode>& node) {
int span = state.end - state.start;
vector<State> new_states;
if (symbols.size() >= max_rule_symbols || state.end + 1 >= word_ids.size() ||
span >= max_rule_span) {
return new_states;
}
new_states.push_back(State(state.start, state.end + 1, symbols,
state.subpatterns_start, node, state.starts_with_x));
int num_subpatterns = phrase.Arity() + state.starts_with_x == 0;
if (symbols.size() + 1 >= max_rule_symbols ||
phrase.Arity() >= max_nonterminals ||
num_subpatterns >= max_chunks) {
return new_states;
}
int var_id = vocabulary->GetNonterminalIndex(phrase.Arity() + 1);
symbols.push_back(var_id);
vector<int> subpatterns_start = state.subpatterns_start;
size_t i = state.end + 1 + min_gap_size;
while (i < word_ids.size() && i - state.start <= max_rule_span) {
subpatterns_start.push_back(i);
new_states.push_back(State(state.start, i, symbols, subpatterns_start,
node->GetChild(var_id), state.starts_with_x));
subpatterns_start.pop_back();
++i;
}
return new_states;
}
|
