File size: 12,631 Bytes
158b61b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 |
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Author: Rico Sennrich
#
# This file is part of moses. Its use is licensed under the GNU Lesser General
# Public License version 2.1 or, at your option, any later version.
"""
Extract syntactic n-grams from dependency treebank in Moses XML format for
training RDLM.
Expected format can be produced with
mosesdecoder/scripts/training/wrapper/conll2mosesxml.py
OOV terminal symbols are mapped to preterminal; OOV nonterminals are mapped
to 0 (<unk>)
"""
from __future__ import print_function, unicode_literals, division
import sys
import codecs
import argparse
# Hack for python2/3 compatibility
from io import open
argparse.open = open
try:
from lxml import etree as ET
except ImportError:
from xml.etree import cElementTree as ET
def create_parser():
parser = argparse.ArgumentParser(
description=(
"Extract syntactic n-grams from parsed corpus in "
"Moses XML format for training RDLM"))
parser.add_argument(
'--input', '-i', type=argparse.FileType('r'), default=sys.stdin,
metavar='PATH',
help='Input file (default: standard input).')
parser.add_argument(
'--output', '-o', type=argparse.FileType('w'), default=sys.stdout,
metavar='PATH',
help='Output file (default: standard output).')
parser.add_argument(
'--mode', type=str, choices=['label', 'head'], required=True,
help='Predict terminals (head) or dependency labels (label).')
parser.add_argument(
'--vocab', metavar='PATH', type=str, required=True,
help=(
"Input layer vocabulary file (one item per line; "
"first line '<unk>')"))
parser.add_argument(
'--output_vocab', metavar='PATH', type=str,
help=(
"Output layer vocabulary file "
"(default: use input layer vocabulary)"))
parser.add_argument(
'--left_context', metavar='INT', type=int, default=3,
help=(
"Size of context vector for left siblings "
"(default: %(default)s)"))
parser.add_argument(
'--right_context', metavar='INT', type=int, default=0,
help=(
"Size of context vector for right siblings "
"(default: %(default)s)"))
parser.add_argument(
'--up_context', metavar='INT', type=int, default=2,
help=(
"Size of context vector for ancestors "
"(default: %(default)s)"))
parser.add_argument(
'--glue_symbol', metavar='STR', type=str, default='Q',
help=(
"Glue symbol. Will be skipped during extraction "
"(default: %(default)s)"))
parser.add_argument(
'--start_symbol', metavar='STR', type=str, default='SSTART',
help=(
"Sentence start symbol. Will be skipped during extraction "
"(default: %(default)s)"))
parser.add_argument(
'--end_symbol', metavar='STR', type=str, default='SEND',
help=(
"Sentence end symbol. Will be skipped during extraction "
"(default: %(default)s)"))
return parser
def escape_text(s):
s = s.replace('|', '|') # factor separator
s = s.replace('[', '[') # syntax non-terminal
s = s.replace(']', ']') # syntax non-terminal
s = s.replace('\'', ''') # xml special character
s = s.replace('"', '"') # xml special character
return s
def get_head(xml):
"""Deterministic heuristic to get head of subtree."""
head = None
preterminal = None
for child in xml:
if not len(child):
preterminal = child.get('label')
head = escape_text(child.text.strip())
return head, preterminal
return head, preterminal
def get_syntactic_ngrams(xml, options, vocab, output_vocab,
parent_heads=None, parent_labels=None):
if len(xml):
# Skip glue rules.
skip_glue_labels = [
options.glue_symbol,
options.start_symbol,
options.end_symbol,
]
if xml.get('label') in skip_glue_labels:
for child in xml:
get_syntactic_ngrams(
child, options, vocab, output_vocab, parent_heads,
parent_labels)
return
# Skip virtual nodes.
skip_virtual_labels = [
'<stop_label>',
'<start_label>',
]
if xml.get('label') in skip_virtual_labels:
return
if not parent_heads:
parent_heads = (
[vocab.get('<root_head>', 0)] * options.up_context)
parent_labels = (
[vocab.get('<root_label>', 0)] * options.up_context)
head, preterminal = get_head(xml)
if not head:
head = '<dummy_head>'
preterminal = head
elif head not in vocab:
head = preterminal
label = xml.get('label')
# syntactic n-gram for root node
int_list = []
int_list.extend([start_head_idx] * options.left_context)
int_list.extend([start_label_idx] * options.left_context)
int_list.extend([stop_head_idx] * options.right_context)
int_list.extend([stop_label_idx] * options.right_context)
int_list.extend(parent_heads)
int_list.extend(parent_labels)
# write root of tree
if options.mode == 'label':
int_list.append(output_vocab.get(label, 0))
options.output.write(' '.join(map(str, int_list)) + '\n')
elif options.mode == 'head' and not head == '<dummy_head>':
int_list.append(vocab.get(label, 0))
int_list.append(
output_vocab.get(head, output_vocab.get(preterminal, 0)))
options.output.write(' '.join(map(str, int_list)) + '\n')
parent_heads.append(vocab.get(head, 0))
parent_labels.append(vocab.get(label, 0))
# virtual start/end-of-subtree tag
if len(xml) > 0:
if options.right_context:
start = ET.Element('tree')
start2 = ET.Element('tree')
start.set('label', '<start_label>')
start2.set('label', 'XY')
start2.text = '<start_head>'
start.append(start2)
xml.insert(0, start)
if options.left_context:
end = ET.Element('tree')
end2 = ET.Element('tree')
end.set('label', '<stop_label>')
end2.set('label', 'XY')
end2.text = '<stop_head>'
end.append(end2)
xml.append(end)
heads = []
preterminals = []
labels = []
for child in xml:
if not len(child):
# Mark that the previous sibling is the head of the
# structure (the head/label are not repeated because they're
# also head/label of the parent).
head_child = '<head_head>'
preterminal_child = head_child
child_label = '<head_label>'
else:
head_child, preterminal_child = get_head(child)
child_label = child.get('label')
if head_child is None:
head_child = '<dummy_head>'
heads.append(head_child)
preterminals.append(preterminal_child)
labels.append(child_label)
heads_idx = [
vocab.get(heads[i], vocab.get(preterminals[i], 0))
for i in range(len(heads))]
labels_idx = [
vocab.get(labels[i], 0)
for i in range(len(labels))]
# Ancestor context is the same for all children.
up_heads = parent_heads[-options.up_context:]
up_labels = parent_labels[-options.up_context:]
skip_special_heads = [
'<dummy_head>',
'<head_head>',
'<stop_head>',
'<start_head>',
]
for i, child in enumerate(xml):
# Skip some special symbols, but recursively extract n-grams
# for its children.
if options.mode == 'head' and heads[i] in skip_special_heads:
parent_heads.append(vocab.get(heads[i], 0))
parent_labels.append(vocab.get(labels[i], 0))
get_syntactic_ngrams(
child, options, vocab, output_vocab, parent_heads,
parent_labels)
parent_heads.pop()
parent_labels.pop()
continue
previous_heads = heads_idx[max(0, i - options.left_context):i]
previous_labels = labels_idx[max(0, i - options.left_context):i]
subsequent_heads = heads_idx[i + 1:i + options.right_context + 1]
subsequent_labels = labels_idx[i + 1:i + options.right_context + 1]
if len(previous_heads) < options.left_context:
previous_heads = (
[start_head_idx] *
(options.left_context - len(previous_heads)) +
previous_heads)
previous_labels = (
[start_label_idx] *
(options.left_context - len(previous_labels)) +
previous_labels)
if len(subsequent_heads) < options.right_context:
subsequent_heads += (
[stop_head_idx] *
(options.right_context - len(subsequent_heads)))
subsequent_labels += (
[stop_label_idx] *
(options.right_context - len(subsequent_labels)))
int_list = []
int_list.extend(previous_heads)
int_list.extend(previous_labels)
int_list.extend(subsequent_heads)
int_list.extend(subsequent_labels)
int_list.extend(up_heads)
int_list.extend(up_labels)
if options.mode == 'label':
int_list.append(output_vocab.get(labels[i], 0))
elif options.mode == 'head':
int_list.append(vocab.get(labels[i], 0))
int_list.append(
output_vocab.get(
heads[i], output_vocab.get(preterminals[i], 0)))
options.output.write(' '.join(map(str, int_list)) + '\n')
parent_heads.append(
vocab.get(heads[i], vocab.get(preterminals[i], 0)))
parent_labels.append(vocab.get(labels[i], 0))
get_syntactic_ngrams(
child, options, vocab, output_vocab, parent_heads,
parent_labels)
parent_heads.pop()
parent_labels.pop()
def load_vocab(path):
v = {}
for i, line in enumerate(open(path, encoding="UTF-8")):
v[line.strip()] = i
return v
def main(options):
vocab = load_vocab(options.vocab)
if options.output_vocab is None:
sys.stderr.write(
"No output vocabulary specified; using input vocabulary.\n")
output_vocab = vocab
else:
output_vocab = load_vocab(options.output_vocab)
global start_head_idx
global start_label_idx
global stop_head_idx
global stop_label_idx
start_head_idx = vocab.get("<start_head>", 0)
start_label_idx = vocab.get("<start_label>", 0)
stop_head_idx = vocab.get("<stop_head>", 0)
stop_label_idx = vocab.get("<stop_label>", 0)
i = 0
for line in options.input:
if i and not i % 50000:
sys.stderr.write('.')
if i and not i % 1000000:
sys.stderr.write('{0}\n'.format(i))
if sys.version_info < (3, 0):
if line == b'\n':
continue
# hack for older moses versions with inconsistent encoding of "|"
line = line.replace(b'&bar;', b'|')
else:
if line == '\n':
continue
# hack for older moses versions with inconsistent encoding of "|"
line = line.replace('&bar;', '|')
xml = ET.fromstring(line)
get_syntactic_ngrams(xml, options, vocab, output_vocab)
i += 1
if __name__ == '__main__':
if sys.version_info < (3, 0):
sys.stderr = codecs.getwriter('UTF-8')(sys.stderr)
sys.stdout = codecs.getwriter('UTF-8')(sys.stdout)
parser = create_parser()
options = parser.parse_args()
main(options)
|