-
Notifications
You must be signed in to change notification settings - Fork 1
/
wordlistlid.py
417 lines (348 loc) · 16.4 KB
/
wordlistlid.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
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
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
"""
A simple approach to Language IDentification (LID).
"""
# Copyright (c) 2012, Constantine Lignos
# All rights reserved.
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions
# are met:
#
# 1. Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in
# the documentation and/or other materials provided with the
# distribution.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
# FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
# COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
# INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
# BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
# LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
# LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
# ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
# POSSIBILITY OF SUCH DAMAGE.
from __future__ import division
import codecs
import re
from random import choice, seed
from string import punctuation
from ConfigParser import RawConfigParser, NoOptionError
import numpy
from lid_constants import UNKNOWN_LANG, NO_LANG, ENGLISH, SPANISH
from lidlists import (SPANISH_TOP32, SPANISH_32PLUS, ENGLISH_TOP32, ENGLISH_32PLUS)
from scalereader import JERBOA_NOTAG
from freqratio import wordlist_ratio, prune_ratios
from codeswitchador import cs_langspresent
# Seed for reproducibility
seed(0)
# Names of our models
MODEL0 = '0.1'
MODEL1 = '1.0'
MODEL1_5 = '1.5'
# Methods for dealing with low conf and unknown words
LOW_METHOD_MLE = "mle"
LOW_METHOD_RANDOM = "random"
LOW_METHOD_UNK = "unk"
LOW_METHODS = (LOW_METHOD_MLE, LOW_METHOD_RANDOM, LOW_METHOD_UNK)
UNK_METHOD_RANDOM = 'random'
UNK_METHOD_LEFT = 'left'
UNK_METHOD_RIGHT = 'right'
UNK_METHODS = (UNK_METHOD_RANDOM, UNK_METHOD_LEFT, UNK_METHOD_RIGHT)
BEST_LOW_METHOD = LOW_METHOD_UNK
BEST_UNK_METHOD = UNK_METHOD_LEFT
# Constants for filtering data
PUNC = set(punctuation)
BAD_TOKENS = set(['rt'])
ALLOWED_TAG_PREFIXES = set(['ja', 'ha'])
class TwoListLID(object):
"""
Provides language identification using two wordlists per language.
"""
def __init__(self, lang_wordlists):
"""
Set up for language identification using the given (lang_name, shortlist, longlist) tuples.
@param lang_wordlists sequence of (lang_name, shortlist, longlist) tuples.
"""
# Force wordlists to be sets
self.langs, self.shorts, self.longs = zip(*[(lang, set(shortlist), set(longlist))
for lang, shortlist, longlist in lang_wordlists])
# For speed, precompute number of langs
self.nlangs = len(self.langs)
def idlangs(self, tokens):
"""
Return whether a language is present and the counts from each wordlist.
@param tokens sequence of tokens to process
@returns array of booleans for language presence and tuples of wordlist scores
The ordering of return vectors will match the order of languages given
at initialization.
"""
# Create scores
short_scores = numpy.zeros(self.nlangs, dtype=numpy.int)
long_scores = numpy.zeros(self.nlangs, dtype=numpy.int)
# Count each language
for token in tokens:
for idx in range(self.nlangs):
if token in self.shorts[idx]:
short_scores[idx] += 1
if token in self.longs[idx]:
long_scores[idx] += 1
# Decide whether each language is there
langspresent = [(short_scores[idx] > 1 or
(short_scores[idx] == 1 and long_scores[idx] > 0))
for idx in range(self.nlangs)]
# Codeswitching verdict
cs = cs_langspresent(langspresent)
# Give the number of hits in the wordlists
hits = zip(short_scores, long_scores)
lid = self._pick_lang(hits)
return (lid, langspresent, hits, cs)
def _pick_lang(self, hits):
"""
Pick the best language from the hits resulting from idlangs using the shortlists.
In case of ties, the first language in self.langs will be the winner.
"""
shortlist_hits, _ = zip(*hits)
if any(shortlist_hits):
return self.langs[numpy.argmax(shortlist_hits)]
else:
return UNKNOWN_LANG
class DefaultTwoListLID(TwoListLID):
"""
A class for doing LID using some boring, untested defaults.
"""
# Order matters, because the first one is the default in ties
LANGS = (SPANISH, ENGLISH)
def __init__(self):
super(DefaultTwoListLID, self).__init__(((SPANISH, SPANISH_TOP32, SPANISH_32PLUS),
(ENGLISH, ENGLISH_TOP32, ENGLISH_32PLUS)))
class RatioListLID(object):
"""Provides LID using a list of frequency ratios for words in two languages."""
UNK_WORD_RATIO = 1.0
_CONFIG_MODEL = 'model'
def __init__(self, ratiodict, lang1, lang2, low_ratio, high_ratio, lang1_min, lang2_min,
lang1_max_unk_rate, lang2_max_unk_rate, cs_max_unk_rate):
"""Set up for language ID using a ratiodict and
@param ratiodict dict of word: ratio pairs
@param lang1 name to give the language whose words have low ratios
@param lang2 name to give the language whose words have high ratios
"""
self._ratios = ratiodict
self.langs = (lang1, lang2, UNKNOWN_LANG)
self.low_ratio = low_ratio
self.high_ratio = high_ratio
self.present_mins = (lang1_min, lang2_min)
self.lang1_max_unk_rate = lang1_max_unk_rate
self.lang2_max_unk_rate = lang2_max_unk_rate
self.cs_max_unk_rate = cs_max_unk_rate
def _ratio_lang(self, ratio):
"""
Return the language for a given ratio.
@param low_ratio ratio below which words need to be to count as lang1
@param high_ratio ratio above which words need to be to count as lang2
"""
if ratio < self.low_ratio:
return self.langs[0]
elif ratio > self.high_ratio:
return self.langs[1]
else:
return UNKNOWN_LANG
def idlangs(self, tokens):
"""
Return whether a language is present and the counts from each wordlist.
@param tokens: tokens to identify
"""
# Per-token ratios and langs
ratios = [self._ratios.get(token, RatioListLID.UNK_WORD_RATIO) for token in tokens]
langs = [self._ratio_lang(ratio) if not non_lid(token) else None
for ratio, token in zip(ratios, tokens)]
# Count hits, making a copy with no UNKNOWN_LANG as well
hits = [langs.count(lang) for lang in self.langs]
known_lang_hits = hits[:-1]
unknown_hits = hits[-1]
hitcount = sum(hits)
unk_rate = unknown_hits / hitcount if hitcount else 1.0
langspresent = [(langhits >= present_min)
for langhits, present_min in zip(known_lang_hits, self.present_mins)]
# Zero out langspresent based on unknown rate
langspresent[0] = langspresent[0] and (unk_rate <= self.lang1_max_unk_rate)
langspresent[1] = langspresent[1] and (unk_rate <= self.lang2_max_unk_rate)
# If we're under the acceptable unknown rate, we can have codeswitching
cs = cs_langspresent(langspresent) if (unk_rate <= self.cs_max_unk_rate) else False
# Compute LID based on the greatest number of hits that passed thresholds
lid = self._pick_lang([hit if present else 0
for hit, present in zip(known_lang_hits, langspresent)])
return (lid, langspresent, hits, ratios, langs, unk_rate, cs)
def _pick_lang(self, hits):
"""
Pick the best language from the hits resulting from idlangs using the shortlists.
In case of ties, the first language in self.langs will be the winner.
"""
if any(hits):
return self.langs[numpy.argmax(hits)]
else:
return UNKNOWN_LANG
@classmethod
def create_from_config(cls, path):
"""Return a RatioListLID class initialized from a config file."""
# Parse configuration from the file
config = RawConfigParser()
result = config.read(path)
if not result:
raise IOError("Couldn't parse config file: %s" % path)
# Read in configuration values
lang1 = config.get(cls._CONFIG_MODEL, "lang1")
lang2 = config.get(cls._CONFIG_MODEL, "lang2")
high_ratio = config.getfloat(cls._CONFIG_MODEL, "high_ratio")
low_ratio = config.getfloat(cls._CONFIG_MODEL, "low_ratio")
wordlist1_path = config.get(cls._CONFIG_MODEL, "wordlist1")
wordlist2_path = config.get(cls._CONFIG_MODEL, "wordlist2")
smoothing = config.getfloat(cls._CONFIG_MODEL, "smoothing")
min_freq = config.getint(cls._CONFIG_MODEL, "min_freq")
lang1_min = config.getint(cls._CONFIG_MODEL, "lang1_min")
lang2_min = config.getint(cls._CONFIG_MODEL, "lang2_min")
lang1_max_unk_rate = config.getfloat(cls._CONFIG_MODEL, "lang1_max_unk_rate")
lang2_max_unk_rate = config.getfloat(cls._CONFIG_MODEL, "lang2_max_unk_rate")
cs_max_unk_rate = config.getfloat(cls._CONFIG_MODEL, "cs_max_unk_rate")
try:
ignorelist = config.get(cls._CONFIG_MODEL, "ignorelist")
except NoOptionError:
ignorelist = None
# Create the structures needed for the LIDder and return it
wordlist1 = codecs.open(wordlist1_path, 'Ur', 'utf_8')
wordlist2 = codecs.open(wordlist2_path, 'Ur', 'utf_8')
ratios, unused1, unused2 = wordlist_ratio(wordlist1, wordlist2, smoothing, min_freq)
# Remove words from the ignorelist
if ignorelist:
bad_words = set(line.strip() for line in codecs.open(ignorelist, 'Ur', 'utf_8'))
prune_ratios(ratios, bad_words)
return cls(ratios, lang1, lang2, low_ratio, high_ratio, lang1_min, lang2_min,
lang1_max_unk_rate, lang2_max_unk_rate, cs_max_unk_rate)
@classmethod
def langs_from_config(cls, path):
"""Return (lang1, lang2) from a config file."""
config = RawConfigParser()
result = config.read(path)
if not result:
raise IOError("Couldn't parse config file: %s" % path)
return (config.get(cls._CONFIG_MODEL, "lang1"), config.get(cls._CONFIG_MODEL, "lang2"))
class AttachingRatioListLID(RatioListLID):
"""Version of RatioListLID that applies unk word attachment as a part of LID/CS."""
def idlangs(self, tokens, lowmethod, unkmethod, tags=None):
"""
Return whether a language is present and the counts from each wordlist.
@param tokens: tokens to identify
@param tags: optional Jerboa tags for the tokens
"""
# Per-token ratios and langs
ratios = [self._ratios.get(token, RatioListLID.UNK_WORD_RATIO) for token in tokens]
langs = [self._ratio_lang(ratio) if not non_lid(token) else None
for ratio, token in zip(ratios, tokens)]
# Put in dummy tags if needed
if not tags:
tags = [JERBOA_NOTAG] * len(tokens)
# Choose langs for
langs = [choose_lang(token, lang, self.langs, tag, ratio, lowmethod, unkmethod, False)
for token, tag, lang, ratio in zip(tokens, tags, langs, ratios)]
# Clean out any remaining unknowns
if None in langs:
langs = choose_unk_lang(langs, unkmethod)
# Count hits, making a copy with no UNKNOWN_LANG as well
hits = [langs.count(lang) for lang in self.langs]
known_lang_hits = hits[:-1]
unknown_hits = hits[-1]
hitcount = sum(hits)
unk_rate = unknown_hits / hitcount if hitcount else 1.0
langspresent = [(langhits >= present_min)
for langhits, present_min in zip(known_lang_hits, self.present_mins)]
# Zero out langspresent based on unknown rate
langspresent[0] = langspresent[0] and (unk_rate <= self.lang1_max_unk_rate)
langspresent[1] = langspresent[1] and (unk_rate <= self.lang2_max_unk_rate)
# If we're under the acceptable unknown rate, we can have codeswitching
cs = cs_langspresent(langspresent) if (unk_rate <= self.cs_max_unk_rate) else False
# Compute LID based on the greatest number of hits that passed thresholds
lid = self._pick_lang([hit if present else 0
for hit, present in zip(known_lang_hits, langspresent)])
return (lid, langspresent, hits, ratios, langs, unk_rate, cs)
def default_lidder(model_name, model_config=None):
"""Return the default lidder for a given model name."""
# Set up the LIDder
if model_name == MODEL0:
return DefaultTwoListLID()
else:
assert model_config, "Must specify model_config for models with parameters"
if model_name == MODEL1:
return RatioListLID.create_from_config(model_config)
elif model_name == MODEL1_5:
return AttachingRatioListLID.create_from_config(model_config)
raise ValueError("Unknown model: %s" % model_name)
ALL_DIGIT_MATCHER = re.compile('^[\.0-9]+$')
def non_lid(token, tag=JERBOA_NOTAG):
"""Return whether a token's LID should be ignored."""
return (all([char in PUNC for char in token]) or
ALL_DIGIT_MATCHER.match(token) or
(tag != JERBOA_NOTAG and tag[:2] not in ALLOWED_TAG_PREFIXES) or
token in BAD_TOKENS)
def choose_lang(token, lang, all_langs, tag, ratio, lowmethod, unkmethod, label_all):
"""Choose what language to assign a token."""
if not label_all and non_lid(token, tag):
# Not a linguistic token
return NO_LANG
elif ratio == RatioListLID.UNK_WORD_RATIO:
# Unknown word
if unkmethod == UNK_METHOD_RANDOM:
return choice(all_langs)
else:
# This will have to be selected later in context
return None
elif lang == UNKNOWN_LANG:
# Low confidence word, but not unknown
if lowmethod == LOW_METHOD_RANDOM:
return choice(all_langs)
elif lowmethod == LOW_METHOD_UNK:
# Leave it for the unknown filter, unless the unknown method is just random
if unkmethod == UNK_METHOD_RANDOM:
return choice(all_langs)
else:
return None
else:
if ratio < 1.0:
return all_langs[0]
else:
return all_langs[1]
else:
return lang
def choose_unk_lang(langs, unkmethod):
"""Replace any remaining unknown langs by rule."""
if unkmethod not in (UNK_METHOD_RIGHT, UNK_METHOD_LEFT):
raise ValueError("Should not be choosing UNK langs when the choice is random.")
# This code is written to go left to right, taking the lang from the left. To go
# from the right, we reverse the list. Either way, we're working on a copy
langs = langs[::-1] if unkmethod == UNK_METHOD_RIGHT else langs[:]
while True:
# Find an unk lang if there is one
try:
unk_idx = langs.index(None)
except ValueError:
break
# Make an ordered list of which tokens it would be a good idea to
# take a lang from. Get everything from the left first going backwards,
# and then everything from the right. If the index is zero, just go right.
next_langs = (langs[unk_idx - 1::-1] + langs[unk_idx + 1:]) if unk_idx else langs[unk_idx + 1:]
for lang in next_langs:
if lang not in (None, NO_LANG):
langs[unk_idx] = lang
break
else:
# Give up and leave it as unknown
langs[unk_idx] = UNKNOWN_LANG
continue
# Ensure we won't loop forever
assert(langs[unk_idx] is not None)
# Reverse again if needed
return langs[::-1] if unkmethod == UNK_METHOD_RIGHT else langs