-
Notifications
You must be signed in to change notification settings - Fork 0
/
phrase_forest.py
251 lines (235 loc) · 8.22 KB
/
phrase_forest.py
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
#!/usr/bin/env python3
import sys
import time
import pickle
import os
import alignment
import common
from common import timed, select
from rule import Rule
from extractor import extract_phrases
import hypergraph
import logger
from phrase_hypergraph import PhraseHGNode, PhraseHGEdge
from alignment import get_reversed_index_map
from consensus_training import cartesian
import rule_filter
from rule_dumper import RuleDumper
import rule_extraction_flags
import gflags
FLAGS = gflags.FLAGS
gflags.DEFINE_string(
'phrase_nonterminal',
'A',
'Nonterminal used for phrase forest.')
gflags.DEFINE_bool(
'delete_unaligned',
False,
'Delete unaligned words in phrase decomposition forest.')
PHRASE_NT = '[%s]' % FLAGS.phrase_nonterminal
def make_rule(parent,
children,
fwords,
ewords,
parent_nt=None,
children_nts=None):
"""Given parent and children as phrases (boxes), return a rule.
A phrase is a list [fi, fj, ei, ej]."""
if parent_nt is None:
parent_nt = PHRASE_NT
if children_nts is None:
children_nts = [PHRASE_NT]*len(children)
fi, fj, ei, ej = parent
f = fwords[fi:fj]
e = ewords[ei:ej]
# maps from index in phrase to index in sentence
# used to keep track of word alignment for lexical weighting
fpos = [i for i in range(fi, fj)]
epos = [i for i in range(ei, ej)]
# None is used as a placeholder in gaps
for var_idx in range(len(children)):
child_fi, child_fj, child_ei, child_ej = children[var_idx]
child_nt = children_nts[var_idx]
# phrase index
sub_fi = child_fi - fi
sub_fj = child_fj - fi
f[sub_fi] = child_nt
fpos[sub_fi] = (child_fi, child_fj)
for i in range(sub_fi+1, sub_fj):
f[i] = None
fpos[i] = None
# phrase index
sub_ei = child_ei - ei
sub_ej = child_ej - ei
e[sub_ei] = (child_nt, var_idx)
epos[sub_ei] = (child_ei, child_ej)
for i in range(sub_ei+1, sub_ej):
e[i] = None
epos[i] = None
# remove placeholders
f = [w for w in f if w is not None]
fpos = [i for i in fpos if i is not None]
epos = [i for i in epos if i is not None]
# recover nonterminal permutation
new_e = []
e2f = []
for w in e:
if w is not None:
if type(w) is tuple:
new_e.append(w[0])
e2f.append(w[1])
else:
new_e.append(w)
# build rule
rule = Rule()
rule.init(parent_nt, f, new_e, e2f)
rule.fpos = fpos
rule.epos = epos
return rule
def phrase_decomposition_forest(align):
# save the index mapping so that we can restore indices after phrase
# decomposition forest generation
if not FLAGS.delete_unaligned:
fmap = get_reversed_index_map(align.faligned)
emap = get_reversed_index_map(align.ealigned)
a = align.remove_unaligned()
phrases = list(extract_phrases(a))
#print('%s phrases' % len(phrases))
n = len(a.fwords)
#print('%s words' % n)
chart = [[None for j in range(n+1)] for i in range(n+1)]
for i1, j1, i2, j2 in phrases:
#print('(%s,%s)' % (i1, i2))
chart[i1][i2] = PhraseHGNode(PHRASE_NT, i1, i2, j1, j2)
for s in range(1, n+1):
for i in range(0, n-s+1):
j = i + s
#print('span (%s %s)' % (i, j))
node = chart[i][j]
if node is None:
continue
splits = 0
# test for binary ambiguity
for k in range(i+1, j):
if chart[i][k] is not None and chart[k][j] is not None:
edge = PhraseHGEdge()
edge.add_tail(chart[i][k])
edge.add_tail(chart[k][j])
node.add_incoming(edge)
#print('split at %s' % k)
splits += 1
# find the maximal cover if no ambiguity found
if splits == 0:
edge = PhraseHGEdge()
l = i
while l < j:
next = l + 1
m = j - 1 if l == i else j
while m > l:
if chart[l][m] is not None:
edge.add_tail(chart[l][m])
next = m
break
m -= 1
l = next
node.add_incoming(edge)
hg = hypergraph.Hypergraph(chart[0][n])
hg.assert_done('topo_sort')
assert len(phrases) == len(hg.nodes), \
'%s phrases, %s nodes' % (len(phrases), len(hg.nodes))
#if len(phrases) != len(hg.nodes):
# print('%s phrases, %s nodes' % (len(phrases), len(hg.nodes)))
#for node in hg.nodes:
# i1,j2,i2,j2 = node.phrase
# print('(%s,%s)' % (i1, i2))
# restore indices on each node
if FLAGS.delete_unaligned:
return hg, a
else:
for node in hg.nodes:
node.fi = max(fmap[node.fi])
node.fj = min(fmap[node.fj])
node.ei = max(emap[node.ei])
node.ej = min(emap[node.ej])
return hg, align
def make_composed_rules(node, align):
node.children = []
node.rules = []
if node.fj - node.fi > 10 or node.ej - node.ei > 10:
return
L = []
for edge in node.incoming:
for children in cartesian([[[tailnode]] + tailnode.children
for tailnode in edge.tail]):
l = []
[l.extend(c) for c in children]
L.append(tuple(l))
# duplicate children set arise from different splits
L = set(L)
for l in L:
rule = make_rule([node.fi, node.fj, node.ei, node.ej],
[[c.fi, c.fj, c.ei, c.ej] for c in l],
align.fwords,
align.ewords)
if rule_filter.filter_box(node, l, align):
rule = make_rule([node.fi, node.fj, node.ei, node.ej],
[[c.fi, c.fj, c.ei, c.ej] for c in l],
align.fwords,
align.ewords)
#if rule_filter.filter(rule):
node.rules.append(rule)
node.children.append(l)
if __name__ == '__main__':
argv = common.parse_flags()
ffilename = FLAGS.parallel_corpus[0]
efilename = FLAGS.parallel_corpus[1]
afilename = FLAGS.parallel_corpus[2]
ffile = open(ffilename)
efile = open(efilename)
afile = open(afilename)
alignments = alignment.Alignment.reader_pharaoh(ffile, efile, afile)
hgs = []
rule_dumper = RuleDumper()
for i, a in enumerate(timed(select(alignments)), 1):
a.write_visual(logger.file)
#if i != 8:
# continue
#logger.writeln('--- %s ---' % i)
#a.write_visual(logger.file)
hg, a = phrase_decomposition_forest(a)
hgs.append(hg)
for node in hg.topo_order():
for edge in node.incoming:
edge.rule = make_rule([edge.head.fi, edge.head.fj, edge.head.ei, edge.head.ej],
[[x.fi, x.fj, x.ei, x.ej] for x in edge.tail],
a.fwords,
a.ewords)
#hg.show()
nnodewithrule = 0
nrules = 0
rules = []
for node in hg.topo_order():
#print('-- node %s ' % node)
#for edge in node.incoming:
#print('- edge %s ' % edge)
#print([len(t.rules) for t in edge.tail])
make_composed_rules(node, a)
#node.make_composed_rules()
if len(node.rules) > 0:
nnodewithrule += 1
#for rule in node.rules:
# print(rule)
#print('%s rules' % len(node.rules))
rules.extend(node.rules)
nrules += len(node.rules)
#print('rules extracted from %s/%s nodes' %
# (nnodewithrule, len(hg.nodes)))
for rule in rules:
print(rule)
rule_dumper.add(rules)
logger.writeln('%s rules extracted from sent %s' % (nrules, i))
#logger.writeln('%s uniq' % len(set(str(r) for r in rules)))
rule_dumper.dump()
ffile.close()
efile.close()
afile.close()