dqnguyen commited on
Commit
359e44d
1 Parent(s): aacc1df

Upload 3 files

Browse files
tokenization_bertweet.py ADDED
@@ -0,0 +1,792 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright (c) 2020, VinAI Research and the HuggingFace Inc. team.
3
+ # Copyright 2018 The Open AI Team Authors and The HuggingFace Inc. team.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+ # limitations under the License.
16
+ """ Tokenization classes for BERTweet"""
17
+
18
+
19
+ import html
20
+ import os
21
+ import re
22
+ from shutil import copyfile
23
+ from typing import List, Optional, Tuple
24
+
25
+ import regex
26
+
27
+ from transformers.tokenization_utils import PreTrainedTokenizer
28
+ from transformers.utils import logging
29
+
30
+
31
+ logger = logging.get_logger(__name__)
32
+
33
+ VOCAB_FILES_NAMES = {
34
+ "vocab_file": "vocab.txt",
35
+ "merges_file": "bpe.codes",
36
+ }
37
+
38
+ PRETRAINED_VOCAB_FILES_MAP = {
39
+ "vocab_file": {
40
+ "vinai/bertweet-base": "https://huggingface.co/vinai/bertweet-base/resolve/main/vocab.txt",
41
+ },
42
+ "merges_file": {
43
+ "vinai/bertweet-base": "https://huggingface.co/vinai/bertweet-base/resolve/main/bpe.codes",
44
+ },
45
+ }
46
+
47
+ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
48
+ "vinai/bertweet-base": 128,
49
+ }
50
+
51
+
52
+ def get_pairs(word):
53
+ """
54
+ Return set of symbol pairs in a word.
55
+
56
+ Word is represented as tuple of symbols (symbols being variable-length strings).
57
+ """
58
+ pairs = set()
59
+ prev_char = word[0]
60
+ for char in word[1:]:
61
+ pairs.add((prev_char, char))
62
+ prev_char = char
63
+
64
+ pairs = set(pairs)
65
+ return pairs
66
+
67
+
68
+ class BertweetTokenizer(PreTrainedTokenizer):
69
+ """
70
+ Constructs a BERTweet tokenizer, using Byte-Pair-Encoding.
71
+
72
+ This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
73
+ this superclass for more information regarding those methods.
74
+
75
+ Args:
76
+ vocab_file (`str`):
77
+ Path to the vocabulary file.
78
+ merges_file (`str`):
79
+ Path to the merges file.
80
+ normalization (`bool`, *optional*, defaults to `False`)
81
+ Whether or not to apply a normalization preprocess.
82
+ bos_token (`str`, *optional*, defaults to `"<s>"`):
83
+ The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
84
+
85
+ <Tip>
86
+
87
+ When building a sequence using special tokens, this is not the token that is used for the beginning of
88
+ sequence. The token used is the `cls_token`.
89
+
90
+ </Tip>
91
+
92
+ eos_token (`str`, *optional*, defaults to `"</s>"`):
93
+ The end of sequence token.
94
+
95
+ <Tip>
96
+
97
+ When building a sequence using special tokens, this is not the token that is used for the end of sequence.
98
+ The token used is the `sep_token`.
99
+
100
+ </Tip>
101
+
102
+ sep_token (`str`, *optional*, defaults to `"</s>"`):
103
+ The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
104
+ sequence classification or for a text and a question for question answering. It is also used as the last
105
+ token of a sequence built with special tokens.
106
+ cls_token (`str`, *optional*, defaults to `"<s>"`):
107
+ The classifier token which is used when doing sequence classification (classification of the whole sequence
108
+ instead of per-token classification). It is the first token of the sequence when built with special tokens.
109
+ unk_token (`str`, *optional*, defaults to `"<unk>"`):
110
+ The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
111
+ token instead.
112
+ pad_token (`str`, *optional*, defaults to `"<pad>"`):
113
+ The token used for padding, for example when batching sequences of different lengths.
114
+ mask_token (`str`, *optional*, defaults to `"<mask>"`):
115
+ The token used for masking values. This is the token used when training this model with masked language
116
+ modeling. This is the token which the model will try to predict.
117
+ """
118
+
119
+ vocab_files_names = VOCAB_FILES_NAMES
120
+ pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
121
+ max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
122
+ model_input_names = ["input_ids", "attention_mask"]
123
+
124
+ def __init__(
125
+ self,
126
+ vocab_file,
127
+ merges_file,
128
+ normalization=False,
129
+ bos_token="<s>",
130
+ eos_token="</s>",
131
+ sep_token="</s>",
132
+ cls_token="<s>",
133
+ unk_token="<unk>",
134
+ pad_token="<pad>",
135
+ mask_token="<mask>",
136
+ **kwargs
137
+ ):
138
+ super().__init__(
139
+ normalization=normalization,
140
+ bos_token=bos_token,
141
+ eos_token=eos_token,
142
+ sep_token=sep_token,
143
+ cls_token=cls_token,
144
+ unk_token=unk_token,
145
+ pad_token=pad_token,
146
+ mask_token=mask_token,
147
+ **kwargs,
148
+ )
149
+
150
+ try:
151
+ from emoji import demojize
152
+
153
+ self.demojizer = demojize
154
+ except ImportError:
155
+ logger.warning(
156
+ "emoji is not installed, thus not converting emoticons or emojis into text. Install emoji: pip3"
157
+ " install emoji==0.6.0"
158
+ )
159
+ self.demojizer = None
160
+
161
+ self.vocab_file = vocab_file
162
+ self.merges_file = merges_file
163
+
164
+ self.encoder = {}
165
+ self.encoder[self.bos_token] = 0
166
+ self.encoder[self.pad_token] = 1
167
+ self.encoder[self.eos_token] = 2
168
+ self.encoder[self.unk_token] = 3
169
+
170
+ self.add_from_file(vocab_file)
171
+ self.encoder[self.mask_token] = len(self.encoder)
172
+
173
+ self.decoder = {v: k for k, v in self.encoder.items()}
174
+
175
+ with open(merges_file, encoding="utf-8") as merges_handle:
176
+ merges = merges_handle.read().split("\n")[:-1]
177
+ merges = [tuple(merge.split()[:-1]) for merge in merges]
178
+ self.bpe_ranks = dict(zip(merges, range(len(merges))))
179
+ self.cache = {}
180
+
181
+ self.normalization = normalization
182
+ self.tweetPreprocessor = TweetTokenizer()
183
+
184
+ self.special_puncts = {"’": "'", "…": "..."}
185
+
186
+ def build_inputs_with_special_tokens(
187
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
188
+ ) -> List[int]:
189
+ """
190
+ Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
191
+ adding special tokens. A BERTweet sequence has the following format:
192
+
193
+ - single sequence: `<s> X </s>`
194
+ - pair of sequences: `<s> A </s></s> B </s>`
195
+
196
+ Args:
197
+ token_ids_0 (`List[int]`):
198
+ List of IDs to which the special tokens will be added.
199
+ token_ids_1 (`List[int]`, *optional*):
200
+ Optional second list of IDs for sequence pairs.
201
+
202
+ Returns:
203
+ `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
204
+ """
205
+
206
+ if token_ids_1 is None:
207
+ return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
208
+ cls = [self.cls_token_id]
209
+ sep = [self.sep_token_id]
210
+ return cls + token_ids_0 + sep + sep + token_ids_1 + sep
211
+
212
+ def get_special_tokens_mask(
213
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
214
+ ) -> List[int]:
215
+ """
216
+ Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
217
+ special tokens using the tokenizer `prepare_for_model` method.
218
+
219
+ Args:
220
+ token_ids_0 (`List[int]`):
221
+ List of IDs.
222
+ token_ids_1 (`List[int]`, *optional*):
223
+ Optional second list of IDs for sequence pairs.
224
+ already_has_special_tokens (`bool`, *optional*, defaults to `False`):
225
+ Whether or not the token list is already formatted with special tokens for the model.
226
+
227
+ Returns:
228
+ `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
229
+ """
230
+
231
+ if already_has_special_tokens:
232
+ return super().get_special_tokens_mask(
233
+ token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
234
+ )
235
+
236
+ if token_ids_1 is None:
237
+ return [1] + ([0] * len(token_ids_0)) + [1]
238
+ return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
239
+
240
+ def create_token_type_ids_from_sequences(
241
+ self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
242
+ ) -> List[int]:
243
+ """
244
+ Create a mask from the two sequences passed to be used in a sequence-pair classification task. BERTweet does
245
+ not make use of token type ids, therefore a list of zeros is returned.
246
+
247
+ Args:
248
+ token_ids_0 (`List[int]`):
249
+ List of IDs.
250
+ token_ids_1 (`List[int]`, *optional*):
251
+ Optional second list of IDs for sequence pairs.
252
+
253
+ Returns:
254
+ `List[int]`: List of zeros.
255
+ """
256
+
257
+ sep = [self.sep_token_id]
258
+ cls = [self.cls_token_id]
259
+
260
+ if token_ids_1 is None:
261
+ return len(cls + token_ids_0 + sep) * [0]
262
+ return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0]
263
+
264
+ @property
265
+ def vocab_size(self):
266
+ return len(self.encoder)
267
+
268
+ def get_vocab(self):
269
+ return dict(self.encoder, **self.added_tokens_encoder)
270
+
271
+ def bpe(self, token):
272
+ if token in self.cache:
273
+ return self.cache[token]
274
+ word = tuple(token)
275
+ word = tuple(list(word[:-1]) + [word[-1] + "</w>"])
276
+ pairs = get_pairs(word)
277
+
278
+ if not pairs:
279
+ return token
280
+
281
+ while True:
282
+ bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
283
+ if bigram not in self.bpe_ranks:
284
+ break
285
+ first, second = bigram
286
+ new_word = []
287
+ i = 0
288
+ while i < len(word):
289
+ try:
290
+ j = word.index(first, i)
291
+ except ValueError:
292
+ new_word.extend(word[i:])
293
+ break
294
+ else:
295
+ new_word.extend(word[i:j])
296
+ i = j
297
+
298
+ if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
299
+ new_word.append(first + second)
300
+ i += 2
301
+ else:
302
+ new_word.append(word[i])
303
+ i += 1
304
+ new_word = tuple(new_word)
305
+ word = new_word
306
+ if len(word) == 1:
307
+ break
308
+ else:
309
+ pairs = get_pairs(word)
310
+ word = "@@ ".join(word)
311
+ word = word[:-4]
312
+ self.cache[token] = word
313
+ return word
314
+
315
+ def _tokenize(self, text):
316
+ """Tokenize a string."""
317
+ if self.normalization: # Perform Tweet normalization before performing BPE
318
+ text = self.normalizeTweet(text)
319
+
320
+ split_tokens = []
321
+ words = re.findall(r"\S+\n?", text)
322
+ for token in words:
323
+ split_tokens.extend([t for t in self.bpe(token).split(" ")])
324
+ return split_tokens
325
+
326
+ def normalizeTweet(self, tweet):
327
+ """
328
+ Normalize a raw Tweet
329
+ """
330
+ for punct in self.special_puncts:
331
+ tweet = tweet.replace(punct, self.special_puncts[punct])
332
+
333
+ tokens = self.tweetPreprocessor.tokenize(tweet)
334
+ normTweet = " ".join([self.normalizeToken(token) for token in tokens])
335
+
336
+ normTweet = (
337
+ normTweet.replace("cannot ", "can not ")
338
+ .replace("n't ", " n't ")
339
+ .replace("n 't ", " n't ")
340
+ .replace("ca n't", "can't")
341
+ .replace("ai n't", "ain't")
342
+ )
343
+ normTweet = (
344
+ normTweet.replace("'m ", " 'm ")
345
+ .replace("'re ", " 're ")
346
+ .replace("'s ", " 's ")
347
+ .replace("'ll ", " 'll ")
348
+ .replace("'d ", " 'd ")
349
+ .replace("'ve ", " 've ")
350
+ )
351
+ normTweet = (
352
+ normTweet.replace(" p . m .", " p.m.")
353
+ .replace(" p . m ", " p.m ")
354
+ .replace(" a . m .", " a.m.")
355
+ .replace(" a . m ", " a.m ")
356
+ )
357
+
358
+ return " ".join(normTweet.split())
359
+
360
+ def normalizeToken(self, token):
361
+ """
362
+ Normalize tokens in a Tweet
363
+ """
364
+ lowercased_token = token.lower()
365
+ if token.startswith("@"):
366
+ return "@USER"
367
+ elif lowercased_token.startswith("http") or lowercased_token.startswith("www"):
368
+ return "HTTPURL"
369
+ elif len(token) == 1:
370
+ if token in self.special_puncts:
371
+ return self.special_puncts[token]
372
+ if self.demojizer is not None:
373
+ return self.demojizer(token)
374
+ else:
375
+ return token
376
+ else:
377
+ return token
378
+
379
+ def _convert_token_to_id(self, token):
380
+ """Converts a token (str) in an id using the vocab."""
381
+ return self.encoder.get(token, self.encoder.get(self.unk_token))
382
+
383
+ def _convert_id_to_token(self, index):
384
+ """Converts an index (integer) in a token (str) using the vocab."""
385
+ return self.decoder.get(index, self.unk_token)
386
+
387
+ def convert_tokens_to_string(self, tokens):
388
+ """Converts a sequence of tokens (string) in a single string."""
389
+ out_string = " ".join(tokens).replace("@@ ", "").strip()
390
+ return out_string
391
+
392
+ def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
393
+ if not os.path.isdir(save_directory):
394
+ logger.error(f"Vocabulary path ({save_directory}) should be a directory.")
395
+ return
396
+
397
+ out_vocab_file = os.path.join(
398
+ save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
399
+ )
400
+
401
+ if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
402
+ copyfile(self.vocab_file, out_vocab_file)
403
+ elif not os.path.isfile(self.vocab_file):
404
+ with open(out_vocab_file, "w", encoding="utf-8") as fp:
405
+ for token, value in self.encoder.items():
406
+ if token not in self.all_special_tokens:
407
+ fp.write(f"{str(token)} 1\n")
408
+
409
+ out_merges_file = os.path.join(
410
+ save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
411
+ )
412
+
413
+ if os.path.abspath(self.merges_file) != os.path.abspath(out_merges_file) and os.path.isfile(self.merges_file):
414
+ copyfile(self.merges_file, out_merges_file)
415
+ elif not os.path.isfile(self.merges_file):
416
+ index = 0
417
+ with open(out_merges_file, "w", encoding="utf-8") as writer:
418
+ for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
419
+ if index != token_index:
420
+ logger.warning(
421
+ f"Saving vocabulary to {out_merges_file}: BPE merge indices are not consecutive."
422
+ " Please check that the tokenizer is not corrupted!"
423
+ )
424
+ index = token_index
425
+ writer.write(" ".join(bpe_tokens) + " 1\n")
426
+ index += 1
427
+
428
+ return (out_vocab_file, out_merges_file)
429
+
430
+ # def decode(self, token_ids, skip_special_tokens=False, clean_up_tokenization_spaces=True):
431
+ # filtered_tokens = ' '.join(self.convert_ids_to_tokens(token_ids, skip_special_tokens=skip_special_tokens))
432
+ # tokens_generated_so_far = re.sub('(@@ )', '', string=filtered_tokens)
433
+ # tokens_generated_so_far = re.sub('(@@ ?$)', '', string=tokens_generated_so_far)
434
+ # return ''.join(tokens_generated_so_far)
435
+
436
+ def add_from_file(self, f):
437
+ """
438
+ Loads a pre-existing dictionary from a text file and adds its symbols to this instance.
439
+ """
440
+ if isinstance(f, str):
441
+ try:
442
+ with open(f, "r", encoding="utf-8") as fd:
443
+ self.add_from_file(fd)
444
+ except FileNotFoundError as fnfe:
445
+ raise fnfe
446
+ except UnicodeError:
447
+ raise Exception(f"Incorrect encoding detected in {f}, please rebuild the dataset")
448
+ return
449
+
450
+ lines = f.readlines()
451
+ for lineTmp in lines:
452
+ line = lineTmp.strip()
453
+ idx = line.rfind(" ")
454
+ if idx == -1:
455
+ raise ValueError("Incorrect dictionary format, expected '<token> <cnt>'")
456
+ word = line[:idx]
457
+ self.encoder[word] = len(self.encoder)
458
+
459
+
460
+ # Natural Language Toolkit: Twitter Tokenizer
461
+ #
462
+ # Copyright (C) 2001-2020 NLTK Project
463
+ # Author: Christopher Potts <[email protected]>
464
+ # Ewan Klein <[email protected]> (modifications)
465
+ # Pierpaolo Pantone <> (modifications)
466
+ # URL: http://nltk.org/
467
+ # For license information, see LICENSE.TXT
468
+ #
469
+
470
+
471
+ """
472
+ Twitter-aware tokenizer, designed to be flexible and easy to adapt to new domains and tasks. The basic logic is this:
473
+
474
+ 1. The tuple regex_strings defines a list of regular expression strings.
475
+
476
+ 2. The regex_strings strings are put, in order, into a compiled regular expression object called word_re.
477
+
478
+ 3. The tokenization is done by word_re.findall(s), where s is the user-supplied string, inside the tokenize() method of
479
+ the class Tokenizer.
480
+
481
+ 4. When instantiating Tokenizer objects, there is a single option: preserve_case. By default, it is set to True. If it
482
+ is set to False, then the tokenizer will lowercase everything except for emoticons.
483
+
484
+ """
485
+
486
+
487
+ ######################################################################
488
+ #
489
+ # import regex # https://github.com/nltk/nltk/issues/2409
490
+ # import html
491
+ #
492
+ ######################################################################
493
+ # The following strings are components in the regular expression
494
+ # that is used for tokenizing. It's important that phone_number
495
+ # appears first in the final regex (since it can contain whitespace).
496
+ # It also could matter that tags comes after emoticons, due to the
497
+ # possibility of having text like
498
+ #
499
+ # <:| and some text >:)
500
+ #
501
+ # Most importantly, the final element should always be last, since it
502
+ # does a last ditch whitespace-based tokenization of whatever is left.
503
+
504
+ # ToDo: Update with http://en.wikipedia.org/wiki/List_of_emoticons ?
505
+
506
+ # This particular element is used in a couple ways, so we define it
507
+ # with a name:
508
+ # docstyle-ignore
509
+ EMOTICONS = r"""
510
+ (?:
511
+ [<>]?
512
+ [:;=8] # eyes
513
+ [\-o\*\']? # optional nose
514
+ [\)\]\(\[dDpP/\:\}\{@\|\\] # mouth
515
+ |
516
+ [\)\]\(\[dDpP/\:\}\{@\|\\] # mouth
517
+ [\-o\*\']? # optional nose
518
+ [:;=8] # eyes
519
+ [<>]?
520
+ |
521
+ <3 # heart
522
+ )"""
523
+
524
+ # URL pattern due to John Gruber, modified by Tom Winzig. See
525
+ # https://gist.github.com/winzig/8894715
526
+ # docstyle-ignore
527
+ URLS = r""" # Capture 1: entire matched URL
528
+ (?:
529
+ https?: # URL protocol and colon
530
+ (?:
531
+ /{1,3} # 1-3 slashes
532
+ | # or
533
+ [a-z0-9%] # Single letter or digit or '%'
534
+ # (Trying not to match e.g. "URI::Escape")
535
+ )
536
+ | # or
537
+ # looks like domain name followed by a slash:
538
+ [a-z0-9.\-]+[.]
539
+ (?:[a-z]{2,13})
540
+ /
541
+ )
542
+ (?: # One or more:
543
+ [^\s()<>{}\[\]]+ # Run of non-space, non-()<>{}[]
544
+ | # or
545
+ \([^\s()]*?\([^\s()]+\)[^\s()]*?\) # balanced parens, one level deep: (...(...)...)
546
+ |
547
+ \([^\s]+?\) # balanced parens, non-recursive: (...)
548
+ )+
549
+ (?: # End with:
550
+ \([^\s()]*?\([^\s()]+\)[^\s()]*?\) # balanced parens, one level deep: (...(...)...)
551
+ |
552
+ \([^\s]+?\) # balanced parens, non-recursive: (...)
553
+ | # or
554
+ [^\s`!()\[\]{};:'".,<>?«»“”‘’] # not a space or one of these punct chars
555
+ )
556
+ | # OR, the following to match naked domains:
557
+ (?:
558
+ (?<!@) # not preceded by a @, avoid matching foo@_gmail.com_
559
+ [a-z0-9]+
560
+ (?:[.\-][a-z0-9]+)*
561
+ [.]
562
+ (?:[a-z]{2,13})
563
+ \b
564
+ /?
565
+ (?!@) # not succeeded by a @,
566
+ # avoid matching "foo.na" in "[email protected]"
567
+ )
568
+ """
569
+
570
+ # docstyle-ignore
571
+ # The components of the tokenizer:
572
+ REGEXPS = (
573
+ URLS,
574
+ # Phone numbers:
575
+ r"""
576
+ (?:
577
+ (?: # (international)
578
+ \+?[01]
579
+ [ *\-.\)]*
580
+ )?
581
+ (?: # (area code)
582
+ [\(]?
583
+ \d{3}
584
+ [ *\-.\)]*
585
+ )?
586
+ \d{3} # exchange
587
+ [ *\-.\)]*
588
+ \d{4} # base
589
+ )""",
590
+ # ASCII Emoticons
591
+ EMOTICONS,
592
+ # HTML tags:
593
+ r"""<[^>\s]+>""",
594
+ # ASCII Arrows
595
+ r"""[\-]+>|<[\-]+""",
596
+ # Twitter username:
597
+ r"""(?:@[\w_]+)""",
598
+ # Twitter hashtags:
599
+ r"""(?:\#+[\w_]+[\w\'_\-]*[\w_]+)""",
600
+ # email addresses
601
+ r"""[\w.+-]+@[\w-]+\.(?:[\w-]\.?)+[\w-]""",
602
+ # docstyle-ignore
603
+ # Remaining word types:
604
+ r"""
605
+ (?:[^\W\d_](?:[^\W\d_]|['\-_])+[^\W\d_]) # Words with apostrophes or dashes.
606
+ |
607
+ (?:[+\-]?\d+[,/.:-]\d+[+\-]?) # Numbers, including fractions, decimals.
608
+ |
609
+ (?:[\w_]+) # Words without apostrophes or dashes.
610
+ |
611
+ (?:\.(?:\s*\.){1,}) # Ellipsis dots.
612
+ |
613
+ (?:\S) # Everything else that isn't whitespace.
614
+ """,
615
+ )
616
+
617
+ ######################################################################
618
+ # This is the core tokenizing regex:
619
+
620
+ WORD_RE = regex.compile(r"""(%s)""" % "|".join(REGEXPS), regex.VERBOSE | regex.I | regex.UNICODE)
621
+
622
+ # WORD_RE performs poorly on these patterns:
623
+ HANG_RE = regex.compile(r"([^a-zA-Z0-9])\1{3,}")
624
+
625
+ # The emoticon string gets its own regex so that we can preserve case for
626
+ # them as needed:
627
+ EMOTICON_RE = regex.compile(EMOTICONS, regex.VERBOSE | regex.I | regex.UNICODE)
628
+
629
+ # These are for regularizing HTML entities to Unicode:
630
+ ENT_RE = regex.compile(r"&(#?(x?))([^&;\s]+);")
631
+
632
+
633
+ ######################################################################
634
+ # Functions for converting html entities
635
+ ######################################################################
636
+
637
+
638
+ def _str_to_unicode(text, encoding=None, errors="strict"):
639
+ if encoding is None:
640
+ encoding = "utf-8"
641
+ if isinstance(text, bytes):
642
+ return text.decode(encoding, errors)
643
+ return text
644
+
645
+
646
+ def _replace_html_entities(text, keep=(), remove_illegal=True, encoding="utf-8"):
647
+ """
648
+ Remove entities from text by converting them to their corresponding unicode character.
649
+
650
+ Args:
651
+ text:
652
+ A unicode string or a byte string encoded in the given *encoding* (which defaults to 'utf-8').
653
+ keep (list):
654
+ List of entity names which should not be replaced. This supports both numeric entities (`&#nnnn;` and
655
+ `&#hhhh;`) and named entities (such as `&nbsp;` or `&gt;`).
656
+ remove_illegal (bool):
657
+ If `True`, entities that can't be converted are removed. Otherwise, entities that can't be converted are
658
+ kept "as is".
659
+
660
+ Returns: A unicode string with the entities removed.
661
+
662
+ See https://github.com/scrapy/w3lib/blob/master/w3lib/html.py
663
+
664
+ >>> from nltk.tokenize.casual import _replace_html_entities >>> _replace_html_entities(b'Price: &pound;100')
665
+ 'Price: \\xa3100' >>> print(_replace_html_entities(b'Price: &pound;100')) Price: £100 >>>
666
+ """
667
+
668
+ def _convert_entity(match):
669
+ entity_body = match.group(3)
670
+ if match.group(1):
671
+ try:
672
+ if match.group(2):
673
+ number = int(entity_body, 16)
674
+ else:
675
+ number = int(entity_body, 10)
676
+ # Numeric character references in the 80-9F range are typically
677
+ # interpreted by browsers as representing the characters mapped
678
+ # to bytes 80-9F in the Windows-1252 encoding. For more info
679
+ # see: https://en.wikipedia.org/wiki/ISO/IEC_8859-1#Similar_character_sets
680
+ if 0x80 <= number <= 0x9F:
681
+ return bytes((number,)).decode("cp1252")
682
+ except ValueError:
683
+ number = None
684
+ else:
685
+ if entity_body in keep:
686
+ return match.group(0)
687
+ else:
688
+ number = html.entities.name2codepoint.get(entity_body)
689
+ if number is not None:
690
+ try:
691
+ return chr(number)
692
+ except (ValueError, OverflowError):
693
+ pass
694
+
695
+ return "" if remove_illegal else match.group(0)
696
+
697
+ return ENT_RE.sub(_convert_entity, _str_to_unicode(text, encoding))
698
+
699
+
700
+ ######################################################################
701
+
702
+
703
+ class TweetTokenizer:
704
+ r"""
705
+ Examples:
706
+
707
+ ```python
708
+ >>> # Tokenizer for tweets.
709
+ >>> from nltk.tokenize import TweetTokenizer
710
+
711
+ >>> tknzr = TweetTokenizer()
712
+ >>> s0 = "This is a cooool #dummysmiley: :-) :-P <3 and some arrows < > -> <--"
713
+ >>> tknzr.tokenize(s0)
714
+ ['This', 'is', 'a', 'cooool', '#dummysmiley', ':', ':-)', ':-P', '<3', 'and', 'some', 'arrows', '<', '>', '->', '<--']
715
+
716
+ >>> # Examples using *strip_handles* and *reduce_len parameters*:
717
+ >>> tknzr = TweetTokenizer(strip_handles=True, reduce_len=True)
718
+ >>> s1 = "@remy: This is waaaaayyyy too much for you!!!!!!"
719
+ >>> tknzr.tokenize(s1)
720
+ [':', 'This', 'is', 'waaayyy', 'too', 'much', 'for', 'you', '!', '!', '!']
721
+ ```"""
722
+
723
+ def __init__(self, preserve_case=True, reduce_len=False, strip_handles=False):
724
+ self.preserve_case = preserve_case
725
+ self.reduce_len = reduce_len
726
+ self.strip_handles = strip_handles
727
+
728
+ def tokenize(self, text):
729
+ """
730
+ Args:
731
+ text: str
732
+
733
+ Returns: list(str) A tokenized list of strings; concatenating this list returns the original string if
734
+ `preserve_case=False`
735
+ """
736
+ # Fix HTML character entities:
737
+ text = _replace_html_entities(text)
738
+ # Remove username handles
739
+ if self.strip_handles:
740
+ text = remove_handles(text)
741
+ # Normalize word lengthening
742
+ if self.reduce_len:
743
+ text = reduce_lengthening(text)
744
+ # Shorten problematic sequences of characters
745
+ safe_text = HANG_RE.sub(r"\1\1\1", text)
746
+ # Tokenize:
747
+ words = WORD_RE.findall(safe_text)
748
+ # Possibly alter the case, but avoid changing emoticons like :D into :d:
749
+ if not self.preserve_case:
750
+ words = list(map((lambda x: x if EMOTICON_RE.search(x) else x.lower()), words))
751
+ return words
752
+
753
+
754
+ ######################################################################
755
+ # Normalization Functions
756
+ ######################################################################
757
+
758
+
759
+ def reduce_lengthening(text):
760
+ """
761
+ Replace repeated character sequences of length 3 or greater with sequences of length 3.
762
+ """
763
+ pattern = regex.compile(r"(.)\1{2,}")
764
+ return pattern.sub(r"\1\1\1", text)
765
+
766
+
767
+ def remove_handles(text):
768
+ """
769
+ Remove Twitter username handles from text.
770
+ """
771
+ pattern = regex.compile(
772
+ r"(?<![A-Za-z0-9_!@#\$%&*])@(([A-Za-z0-9_]){20}(?!@))|(?<![A-Za-z0-9_!@#\$%&*])@(([A-Za-z0-9_]){1,19})(?![A-Za-z0-9_]*@)"
773
+ )
774
+ # Substitute handles with ' ' to ensure that text on either side of removed handles are tokenized correctly
775
+ return pattern.sub(" ", text)
776
+
777
+
778
+ ######################################################################
779
+ # Tokenization Function
780
+ ######################################################################
781
+
782
+
783
+ def casual_tokenize(text, preserve_case=True, reduce_len=False, strip_handles=False):
784
+ """
785
+ Convenience function for wrapping the tokenizer.
786
+ """
787
+ return TweetTokenizer(preserve_case=preserve_case, reduce_len=reduce_len, strip_handles=strip_handles).tokenize(
788
+ text
789
+ )
790
+
791
+
792
+ ###############################################################################
tokenization_bertweet_fast.py CHANGED
@@ -24,7 +24,7 @@ from transformers.tokenization_utils_base import EncodingFast
24
 
25
  from transformers.tokenization_utils_fast import PreTrainedTokenizerFast
26
  from transformers.utils import logging
27
- from transformers import BertweetTokenizer
28
 
29
 
30
  logger = logging.get_logger(__name__)
 
24
 
25
  from transformers.tokenization_utils_fast import PreTrainedTokenizerFast
26
  from transformers.utils import logging
27
+ from .tokenization_bertweet import BertweetTokenizer
28
 
29
 
30
  logger = logging.get_logger(__name__)