forked from gouwsmeister/TextCleanser
-
Notifications
You must be signed in to change notification settings - Fork 0
/
generator.py
566 lines (488 loc) · 23.2 KB
/
generator.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
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
"""
This module generates candidate words by comparing possibly noisy
words against words in the lexicon and returning the words it considers
most likely candidates.
Author: Stephan Gouws
Contact: [email protected]
"""
import nltk
import collections, itertools
import shelve, cPickle
import heapq
import time, math
import re, string, csv
from phonetic_algorithms import PhoneticAlgorithms
from string_functions import StringFunctions
# constants
LEXICON_FILENAME = "data/combined_lex.pickl"
PHON_LEX_FILENAME = "data/phonetic_lex.pickl"
PHONE_LEX_KEYS_FILENAME = "data/phonetic_lex_keys.pickl"
SUB_LEX_FILENAME = "data/sub_lexicon.pickl"
STOPWORDS = cPickle.load(open("data/stopwords-set.pickl", "rb")) #set(nltk.corpus.stopwords.words())
ALPHABET = 'abcdefghijklmnopqrstuvwxyz'
CONSONANTS = "bcdfghjklmnpqrstvwxyz"
PUNC = string.punctuation
#LATTICE_TOOL_DIR = "/media/Data/work/Tools/srilm/bin/i686-m64/"
EMPTY_SYM = "EMPTYSYM" # placeholder for empty symbol, such that tokeniser doesn't split it
class Generator:
# CLASS CONSTANTS
# Candidate word selection methods
IBM_SIM=0
SSK_SIM=1
PHONETIC_ED_SIM=2
def __init__(self, lexicon=None):
# load the lexicon from a pickled dump
# we use the Top20K words in gigaword intersected with CMU
# pronunciation dictionary, adapt this to your needs
self.lexicon = cPickle.load(open(LEXICON_FILENAME, 'rb'))
# load the phonetically indexed lexicon
try:
self.phon_lex = cPickle.load(open(PHON_LEX_FILENAME, 'rb'))
# print "Loaded phonetically indexed lexicon with %d partitions." % (len(self.phon_lex.keys()))
except IOError:
# build phonetic lexicon
self.phon_lex = self.build_phonetic_index()
cPickle.dump(self.phon_lex, open(PHON_LEX_FILENAME, 'wb'))
# load the list of phonetic keys
self.phonetic_keys = self.phon_lex.keys()
# load lexicon sub-indexed by first consonant
try:
self.sub_lexicon = cPickle.load(open(SUB_LEX_FILENAME, 'rb'))
# print "Loaded sub-indexed lexicon with %d partitions." % (len(self.sub_lexicon.keys()))
except IOError:
# build sub-lexicon
# self.lexicon = cPickle.load(open(LEXICON_FILENAME, 'rb'))
#for w in self.lexicon: # remove one-letter entries
# if len(w)<=1:
# del w
self.sub_lexicon = self.build_sub_index()
cPickle.dump(self.sub_lexicon, open(SUB_LEX_FILENAME, 'wb'))
# load list of common abbreviations, adapt common_abbrs.csv to your own needs
self.abbr_list = {}
for row in csv.reader(open("data/common_abbrs.csv"), delimiter=","):
self.abbr_list[row[0]] = row[1]
# print "Loaded list of %d common abbreviations." % len(self.abbr_list)
# initialise translit dictionary
translit_inits = {'0': ['0','o'],
'1':['1','l','one'],
'2':['2','to', 'two'],
'3':['3','e', 'three'],
'4':['4','a', 'four', 'for'],
'5':['5','s', 'five'],
'6':['6','six','b'],
'7':['7','t', 'seven'],
'8':['8','eight','ate'],
'9':['9','nine','g'],
'@':['@', 'at'],
'&':['&', 'and']}
self.translit = {}
self.translit.update(translit_inits)
# print "Loaded lexicon of size: ", len(self.lexicon)
# generate K possible noisy expansions per word.
self.topK = 10
# load phonetic similarity function
self.phon_sim = PhoneticAlgorithms().double_metaphone
# load subsequence-kernel similarity function
ssk=StringFunctions(lamb=0.8, p=2)
self.ssk_sim = ssk.SSK
# smiley regex
self.smiley_regex = re.compile("[:;]-?[DP()\\\|bpoO0]{1,2}") # recognise emoticons
self.lt_gt_regex = re.compile("\&[lr]t;[\d]?")
#pnc = ''.join(l for l in [string.punctuation]) + " \t" # remove double punctuation and spaces
# regexes for detecting usernames, hashtags, rt's and urls at the token level
self.hashTags = re.compile("^#(.*)")
self.username = re.compile("@\w+")
self.rt = re.compile("^rt")
# this regex from: http://snipplr.com/view/36992/improvement-of-url-interpretation-with-regex/
# I made the http required (+) since otherwise it matches things like 'River.With'
# where users did not put a space between the end of sentence and next word
self.urls = re.compile("((https?://)+([-\w]+\.[-\w\.]+)+\w(:\d+)?(/([-\w/_\.]*(\?\S+)?)?)*)")
#self.sim_cache = shelve.open('sim_cache')
def build_phonetic_index(self):
"""Index the lexicon phonetically to reduce lookup time. However this failed
epically and only reduced search space by half on average..."""
print "Building phonetic index over %d words." % (len(self.lexicon))
t1 = time.time()
phon_lex = collections.defaultdict(list)
dbl_metaphone = PhoneticAlgorithms().doubleMetaphone
for w in self.lexicon:
m = dbl_metaphone(w)
phon_lex[m[0]].append(w)
if m[1] != None:
phon_lex[m[1]].append(w)
print "Finished building phonetic index in %d seconds." % (time.time() - t1)
return phon_lex
def build_sub_index(self):
"""Index the lexicon by the first consonant, or first letter if no
consonant found, for faster lookup"""
cons_lex = collections.defaultdict(list)
for w in self.lexicon:
first_letter = self.get_first_cons(w)
cons_lex[first_letter].append(w)
return cons_lex
#@profile
def lcs_len1(self, xs, ys):
'''
from: http://wordaligned.org/articles/longest-common-subsequence
Return the length of the LCS of xs and ys.
Example:
>>> lcs_length("HUMAN", "CHIMPANZEE")
4
'''
ny = len(ys)
curr = list(itertools.repeat(0, 1 + ny))
for x in xs:
prev = list(curr)
for i, y in enumerate(ys):
if x == y:
curr[i+1] = prev[i] + 1
else:
curr[i+1] = max(curr[i], prev[i+1])
return curr[ny]
#@profile
def lcs_len2(self, X, Y):
'''
http://en.wikibooks.org/wiki/Algorithm_implementation/Strings/Longest_common_subsequence#Computing_the_length_of_the_LCS
'''
m = len(X)
n = len(Y)
# An (m+1) times (n+1) matrix
C = [[0] * (n+1) for i in range(m+1)]
for i in range(1, m+1):
for j in range(1, n+1):
if X[i-1] == Y[j-1]:
C[i][j] = C[i-1][j-1] + 1
else:
C[i][j] = max(C[i][j-1], C[i-1][j])
return C[i][j]
#@profile
def cs(self, s):
'''
generates the consonant skeleton of a word, 'shop' -> 'shp'
'''
return ''.join([l for l in s.lower() if l not in 'aeiou'])
#@profile
def edit_dist(self, s1, s2):
if len(s1) < len(s2):
return self.edit_dist(s2, s1)
if not s1:
return len(s2)
previous_row = xrange(len(s2) + 1)
for i, c1 in enumerate(s1):
current_row = [i + 1]
for j, c2 in enumerate(s2):
insertions = previous_row[j + 1] + 1 # j+1 instead of j since previous_row and current_row are one character longer
deletions = current_row[j] + 1 # than s2
substitutions = previous_row[j] + (c1 != c2)
current_row.append(min(insertions, deletions, substitutions))
previous_row = current_row
return previous_row[-1]
#@profile
def lcs_ratio(self, word1, word2):
'''
Return the LCS / max(len(w1),len(w2))
'''
return float(self.lcs_len1(word1,word2)) / max(len(word1),len(word2))
#@profile
def contractor_sim(self, s1, s2):
'''
Implementation of Contractor et al.'s similarity function
'''
"""# Cache results
key = s1+'.'+s2
key = key.encode("utf-8")
sim_val = 0.0
if key not in self.sim_cache:
lcs_ratio = self.lcs_ratio(s1,s2)
sim_val = self.sim_cache[key] = float(lcs_ratio / (self.edit_dist(self.cs(s1), self.cs(s2))+1))
else:
sim_val = self.sim_cache[key]"""
lcs_ratio = self.lcs_ratio(s1,s2)
sim_val = float(lcs_ratio / (self.edit_dist(self.cs(s1), self.cs(s2))+1))
return sim_val
def phonetic_ed_sim(self, s1, s2):
"""Computes the phonetic edit distance SIMILARITY between two strings"""
return math.exp(-(self.edit_dist(s1, s2)))
#@profile
def expand_word(self, noisyWord):
'''
Expand 1) transliterate letters 2) ??...
'''
expTmp = [''] # .. and empty candidate strings
# generate possible transliterations
for l in noisyWord:
expansions = []
try:
translits = self.translit[l]
except KeyError:
translits = [l]
for e in expTmp:
for c in translits:
expansions.append(e + c)
expTmp = expansions[:]
return expTmp #.append(self.expand_token(noisyWord))
def get_first_cons(self, word):
"""Get first consonant or just first letter if no consonant found"""
for l in word:
if l in CONSONANTS:
return l
return word[0]
def word_edits(self, word):
"""Norvig's spelling corrector code for generating off-by-one candidates"""
splits = [(word[:i], word[i:]) for i in range(len(word) + 1)]
deletes = [a + b[1:] for a, b in splits if b]
transposes = [a + b[1] + b[0] + b[2:] for a, b in splits if len(b)>1]
replaces = [a + c + b[1:] for a, b in splits for c in ALPHABET if b]
inserts = [a + c + b for a, b in splits for c in ALPHABET]
return set(deletes + transposes + replaces + inserts)
def expand_abbrs(self, tok):
"""Look for and expand commonly occurring Internet and other abbreviations.
Edit the list in data/common_abbrs.csv.
Note that it is a deterministic replacement.
"""
try:
return self.abbr_list[tok]
except KeyError:
return None
def rank_candidates(self, candidates, lexicon_list, sim_function, gen_off_by_ones):
"""Rank all tokens in candidates with tokens found in lexicon. To make the use
of sub-lexicons possible, it uses lexicon_list[i] for i,token in enumerate(candidates),
i.e. you pass a lexicon_list[i] for each token
TODO: Might use lexicon and sub_lex parameters instead?
"""
# Generate all 'off-by-one' variants
if gen_off_by_ones:
off_by_ones = []
for candidate in candidates:
off_by_ones.extend([w for w in self.word_edits(candidate)]) # if w in lexicon]
candidates.extend(list(set(off_by_ones)))
#print len(noisyWords), " after adding edits."
# compute scores with words in lexicon
K = self.topK
topK = [] # top K heap of candidate words
curW = 0
for i,cand_token in enumerate(candidates):
#for w in self.lexicon:
lexicon=lexicon_list[i]
for w in lexicon:
#if len(w)>=len(cand_token):
sim = sim_function(cand_token, w)
if sim==0: # don't add zero-prob words
continue
if curW < K: # first K, just insert into heap
heapq.heappush(topK, (sim, w))
else: # next (N-K), insert if sim>smallest sim in heap (root)
try:
if (sim > topK[0][0]):
heapq.heapreplace(topK, (sim,w))
except IndexError:
heapq.heappush(topK, (sim, w))
curW += 1
#return heapq.nlargest(K, topK, key=lambda x:x[0])
conf_set = heapq.nlargest(K-1, topK, key=lambda x:x[0])
# remove all 0-prob elements
for elem in conf_set:
if elem[0]==0:
del elem
return conf_set
def hash_user_rt_url(self, tok):
#if noisyWord in ['rt', 'hshtg', 'usr', 'url', EMPTY_SYM]:
hashTags=self.hashTags
username=self.username
rt=self.rt
urls=self.urls
if hashTags.search(tok) or username.search(tok) or rt.search(tok) or urls.search(tok):
return True
else:
return False
def check_oov_but_valid(self, noisyWord):
# check if it is a common abbreviation or contracted form
# TODO: Is this the right place to do this?
abbr = self.expand_abbrs(noisyWord.strip("'"))
if abbr != None:
# TODO: this never replaces lol, etc with EMPTY_SYM after decoding
# return [(0.8, abbr), (0.2, noisyWord)]
return [(1.0, abbr)]
# check if it's a number, longer than 1 char
# TODO: Do better detection to catch any token with no letters
# e.g. "9/11" (dates), "4-4" (scores), etc. and ignore
# Simple first idea, if it contains _no letters_, ignore
if not re.search("[a-z]", noisyWord):
return [(1.0, noisyWord)]
# check for punctuation
if noisyWord in PUNC:
return [(1.0, noisyWord)]
# check for USR or URL or EMPTY_SYM special tokens
# noisy=hashTags.sub('hshtg', username.sub('usr', rt.sub('rt', urls.sub('url', noisy))))
if self.hash_user_rt_url(noisyWord) or noisyWord in [EMPTY_SYM]:
return [(1.0, noisyWord)]
# otherwise
return None
#@profile
def word_generate_candidates(self, noisyWord, rank_method, off_by_ones=False):
"""Generate a confusion set of possible candidates for a word using some rank_method,
currently supported methods include:
Generator.IBM_SIM - implementation of the heuristic used in Contractor et al. 2010
Generator.SSK_SIM - a 2-char string subsequence similarity function
Generator.PHONET_ED_SIM - a phonetic edit distance"""
oov_but_valid = self.check_oov_but_valid(noisyWord)
if oov_but_valid:
return oov_but_valid
# TODO: This seems clumsy, lexicon should include STOPWORDS by default
if noisyWord not in self.lexicon and noisyWord not in STOPWORDS: # O(1) for sets
# expand noisy word
noisyWords = self.expand_word(noisyWord)
# 1) IBM_SIMILARITY and SUBSEQUENCE-KERNEL SIMILARITY
if rank_method in [Generator.IBM_SIM, Generator.SSK_SIM]:
first_letters = [self.get_first_cons(w) for w in noisyWords]
lexicon=[self.sub_lexicon[first_letter] for first_letter in first_letters]
if rank_method == Generator.IBM_SIM:
sim_function=self.contractor_sim
else:
sim_function=self.ssk_sim
candidates=noisyWords
conf_set = self.rank_candidates(candidates, lexicon, sim_function, off_by_ones)
# 2) PHONETIC EDIT DISTANCE SIMILARITY
elif rank_method==Generator.PHONETIC_ED_SIM:
# candidates=set([self.phon_sim(c)[0] for c in noisyWords])
# Include both primary and secondary codes!
candidates=[]
for c in noisyWords:
phonetic_c = self.phon_sim(c)
if phonetic_c[0]:
candidates.append(phonetic_c[0])
if phonetic_c[1]:
candidates.append(phonetic_c[1])
# sim_function=self.edit_dist
lexicon=[self.phonetic_keys for i in range(len(candidates))]
sim_function=self.phonetic_ed_sim
phon_conf_set = self.rank_candidates(candidates, lexicon, sim_function, off_by_ones)
# print "noisyWords: ", noisyWords
# print "phonetic codes: ", candidates
# print "phonetic confusion set: ", phon_conf_set
# expand phonetic codes into their likely candidate words
conf_set = []
for sim,phonetic_code in phon_conf_set:
conf_set.extend([(sim,w) for w in self.phon_lex[phonetic_code]])
# retain the top 20
conf_set = conf_set[:10]
else:
raise NotImplementedError("Unknown rank_method supplied as argument.")
# heuristic: add original word with same prob as lowest prob word in case it's a valid OOV word!
try:
weight = conf_set[-1][0]
except IndexError:
weight = 1.0 # there's nothing in the confusion set
if weight==0:
weight = 0.2 # add original word with some low probability
conf_set.append((weight, noisyWord))
return conf_set
else:
# this is a valid word in the lexicon
# NOTE: This would miss (i) (special case) accidental homophilous misspellings, such as 'rite' and 'right'
# also (ii) (the general case) spelling errors where a spelling error leads to a valid word in the lexicon
# 'it' -> 'is', 'are' -> 'art', 'party' -> 'part', etc.
# only way to get (ii) is to use a spelling error approach and include all 'off-by-one' errors in the
# confusion set, e.g. as computed by Norvig's spelling corrector. For (i) need to use phonetic lookup.
return [(1.0, noisyWord)]
def get_OOV(self, noisyWord):
if self.check_oov_but_valid(noisyWord) or noisyWord in self.lexicon or noisyWord in STOPWORDS:
return []
else:
return self.expand_word(noisyWord)
def sent_preprocess(self, sent):
# collapse punctuation marks occurring > 1 into one
#rem_dbl_punc = re.compile(r"([%s])\1+" % (pnc)) # remove punctuation chars repeated > once
pass
def fix_bad_tokenisation(self, tokens):
"""Fix some general tokenisation issues"""
# Split run-on tokens where we have two words separated by a punctuation mark,
# or one word with leading or trailing punctuation
punc = [p for p in PUNC if p not in ("'", "@", "#")]
split_runon_tokens = re.compile("(\w*)([%s]*)(\w*)" % (punc))
out_tokens = []
for tok in tokens:
if self.hash_user_rt_url(tok):
# keep as is so we don't split '@user'->['@', 'user'], or urls into god knows what
out_tokens.append(tok)
else:
out_tokens.extend([tok for tok in split_runon_tokens.sub(r'\1 \2 \3', tok).split(' ') if tok!=''])
return out_tokens
def sent_generate_candidates(self, sent, rank_method, off_by_ones=False, log_OOV=False):
"""Generate 'confusion set' from a sentence.
Return (r,w,c) replacements made (smileys), words after tokenisation and confusion set."""
# perform some simple pre-processing
# 1) remove emoticons
# for bookkeeping, keep track of all replacements
replacements = []
# For the evaluation task, do not remove smileys
"""for change in self.smiley_regex.findall(sent):
replacements.append((change, '')) # (x,y) for x->y
# TODO: Handle 'left-handed' emoticons..
sent = self.smiley_regex.sub('', sent)
"""
# replace all spurious < and > html artifacts..
sent = self.lt_gt_regex.sub('', sent)
# split into words
# TODO: Get a better tokeniser...
words = sent.split()
# fix bad tokenisation issues
words = self.fix_bad_tokenisation(words)
if log_OOV:
confusion_set=[self.get_OOV(nw.lower()) for nw in words]
else:
# get candidates for each word
confusion_set = [self.word_generate_candidates(nw.lower(), rank_method, off_by_ones) for nw in words]
return replacements, words, confusion_set
def generate_word_mesh(self, confusion_net):
"""
A word mesh is a more constrained form of a word lattice.
It follows much more readily given the confusion network input,
and lets SRI-LM create the lattice itself.
"""
conf_net = []
for nodes in confusion_net:
# remove zero probability words and empty words
keepnodes = [n for n in nodes if n[0]>0 and n[1]!='']
if len(keepnodes)>0:
conf_net.append(keepnodes)
else:
# print "NODES: ", nodes
conf_net.append([(1.0, '*E*')])
pfsg_out = ["name SOMENAME\n"]
pfsg_out.append("numaligns " + str(len(conf_net)) + "\n")
pfsg_out.append("posterior 1\n")
for i,nodes in enumerate(conf_net):
align_str = "align " + str(i)
# normalising constant for nodes' probabilities
# print to_nodes
nodes_total = sum([n[0] for n in nodes])
for node in nodes:
p=node[0]
w=node[1]
align_str += " %s %f" % (w, p/nodes_total)
align_str += "\n"
pfsg_out.append(align_str)
return ''.join(pfsg_out)
"""
TODO ideas list:
1) pre-process and normalise letters occurring more than 2 consecutive times..
2) Implement a recaser, using the original sentence as guide for recasing clean sentence
3) Find suitable heuristic to avoid 'normalising' words which are correct e.g.
neologisms, names, etc. This is tougher and might require training a classifier to
detect OOV words, as in (Han and Baldwin, 2011).
"""
if __name__ == "__main__":
testWords = ['t0day', 'today', 't0d4y', 'w0t', 'wh4t']
testSents = ['test sentence one', 'test s3ntens tw0', 't0day iz awssam']
test_confusion_set = [[(0.4, "w1"), (0.6, "w2")], [(0.3, "w3"),(0.3, "w4"), (0.4, "w5")]]
gen = Generator()
for w in testWords:
# print gen.expand_word(w)
print gen.word_generate_candidates(w)
for s in testSents:
c = gen.sent_generate_candidates(s)
print "Sentence: ", s
print "Candidate list: ", c
word_lattice = gen.generate_word_mesh(c)
print "Word mesh: ", word_mesh