Datasets:

Languages:
English
Multilinguality:
multilingual
Size Categories:
100M<n<1B
Language Creators:
found
Annotations Creators:
no-annotation
Source Datasets:
original
ArXiv:
License:
system HF staff commited on
Commit
213288b
0 Parent(s):

Update files from the datasets library (from 1.0.0)

Browse files

Release notes: https://github.com/huggingface/datasets/releases/tag/1.0.0

Files changed (3) hide show
  1. .gitattributes +27 -0
  2. c4.py +331 -0
  3. c4_utils.py +489 -0
.gitattributes ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bin.* filter=lfs diff=lfs merge=lfs -text
5
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.model filter=lfs diff=lfs merge=lfs -text
12
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
13
+ *.onnx filter=lfs diff=lfs merge=lfs -text
14
+ *.ot filter=lfs diff=lfs merge=lfs -text
15
+ *.parquet filter=lfs diff=lfs merge=lfs -text
16
+ *.pb filter=lfs diff=lfs merge=lfs -text
17
+ *.pt filter=lfs diff=lfs merge=lfs -text
18
+ *.pth filter=lfs diff=lfs merge=lfs -text
19
+ *.rar filter=lfs diff=lfs merge=lfs -text
20
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
21
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
22
+ *.tflite filter=lfs diff=lfs merge=lfs -text
23
+ *.tgz filter=lfs diff=lfs merge=lfs -text
24
+ *.xz filter=lfs diff=lfs merge=lfs -text
25
+ *.zip filter=lfs diff=lfs merge=lfs -text
26
+ *.zstandard filter=lfs diff=lfs merge=lfs -text
27
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
c4.py ADDED
@@ -0,0 +1,331 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2020 The TensorFlow Datasets Authors and the HuggingFace Datasets Authors.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ # Lint as: python3
17
+ """C4 dataset based on Common Crawl."""
18
+
19
+ from __future__ import absolute_import, division, print_function
20
+
21
+ import json
22
+ import logging
23
+ import os
24
+
25
+ import datasets
26
+
27
+ from .c4_utils import (
28
+ dedupe_urls,
29
+ filter_by_webtextlike,
30
+ get_clean_page_fn,
31
+ get_counter_inc_fn,
32
+ get_hashed_url_filter_fn,
33
+ is_language,
34
+ is_realnews_domain,
35
+ is_valid_length,
36
+ normalize_url,
37
+ remove_duplicate_text,
38
+ split_wet_file,
39
+ )
40
+
41
+
42
+ _DESCRIPTION = """\
43
+ A colossal, cleaned version of Common Crawl's web crawl corpus.
44
+
45
+ Based on Common Crawl dataset: "https://commoncrawl.org"
46
+
47
+ Due to the overhead of cleaning the dataset, it is recommend you prepare it with
48
+ a distributed service like Cloud Dataflow. More info at
49
+ https://www.tensorflow.org/datasets/beam_datasets.
50
+ """
51
+ _CITATION = """
52
+ @article{2019t5,
53
+ author = {Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu},
54
+ title = {Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer},
55
+ journal = {arXiv e-prints},
56
+ year = {2019},
57
+ archivePrefix = {arXiv},
58
+ eprint = {1910.10683},
59
+ }
60
+ """
61
+ _VERSION = datasets.Version("2.3.0", "Deduplicate lines within a page.")
62
+
63
+ _DOWNLOAD_HOST = "https://commoncrawl.s3.amazonaws.com"
64
+ _WET_PATH_URL = "https://commoncrawl.s3.amazonaws.com/crawl-data/CC-MAIN-{cc_version}/wet.paths.gz"
65
+ _REALNEWS_DOMAINS_URL = "https://raw.githubusercontent.com/rowanz/grover/38f7184bd87237ae2d3bc330b99f1e2e246f6d51/realnews/domain_to_allowed_subdomains.json"
66
+ _BADWORDS_URL = "https://raw.githubusercontent.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words/25e679f03d96baa721cde20db9944649e8d0a844/{lang}"
67
+ _CHECKSUMS_URL = "https://storage.googleapis.com/tfds-data/manual_checksums/c4.txt"
68
+ _OPENWEBTEXT_URLS_ZIP = "OpenWebText.zip"
69
+ _OPENWEBTEXT_URLS_URL = "https://mega.nz/#F!EZZD0YwJ!9_PlEQzdMVLaNdKv_ICNVQ"
70
+ _OPENWEBTEXT_URLS_FILE_PATTERN = "OpenWebText/Version 1/URLs/*.txt"
71
+
72
+ _DEFAULT_CC_VERSIONS = ("2019-18",) # April 2019
73
+ _DEFAULT_WEBTEXTLIKE_CC_VERSIONS = ( # August 2018 - July 2019
74
+ "2018-34",
75
+ "2018-39",
76
+ "2018-43",
77
+ "2018-47",
78
+ "2018-51",
79
+ "2019-04",
80
+ "2019-09",
81
+ "2019-13",
82
+ "2019-18",
83
+ "2019-22",
84
+ "2019-26",
85
+ "2019-30",
86
+ )
87
+
88
+
89
+ class C4Config(datasets.BuilderConfig):
90
+ """BuilderConfig for C4 dataset."""
91
+
92
+ def __init__(self, language, cc_versions=None, clean=True, realnewslike=False, webtextlike=False, **kwargs):
93
+ """BuilderConfig for C4.
94
+
95
+ Args:
96
+ language: string, the language code, or "all" to disable language
97
+ filtering.
98
+ cc_versions: tuple(string), a collection of versions of Common Crawl to
99
+ use as the raw source text. Set to None to use defaults.
100
+ clean: bool, whether to clean the dataset for badwords, duplications, etc.
101
+ realnewslike: bool, whether to limit to news domains as compiled by
102
+ RealNews.
103
+ webtextlike: bool, whether to limit to WebText-like URLs.
104
+ **kwargs: keyword arguments forwarded to super.
105
+ """
106
+ name_parts = [language]
107
+ if cc_versions:
108
+ name_parts.append("_".join(cc_versions))
109
+ if not clean:
110
+ name_parts.append("noclean")
111
+ if realnewslike:
112
+ name_parts.append("realnewslike")
113
+ if webtextlike:
114
+ name_parts.append("webtextlike")
115
+ name = ".".join(name_parts)
116
+ super(C4Config, self).__init__(name=name, version=_VERSION, **kwargs)
117
+ self.lang = language
118
+ self.cc_versions = cc_versions or (_DEFAULT_WEBTEXTLIKE_CC_VERSIONS if webtextlike else _DEFAULT_CC_VERSIONS)
119
+ self.clean = clean
120
+ self.realnewslike = realnewslike
121
+ self.webtextlike = webtextlike
122
+
123
+
124
+ class C4(datasets.BeamBasedBuilder):
125
+ """C4 dataset based on Common Crawl."""
126
+
127
+ BUILDER_CONFIGS = [
128
+ C4Config(language="en", description="English C4 dataset."),
129
+ C4Config(
130
+ language="en",
131
+ clean=False,
132
+ description="Disables all cleaning (deduplication, removal based on bad words, " "etc.)",
133
+ ),
134
+ C4Config(
135
+ language="en",
136
+ realnewslike=True,
137
+ description="Filters from the default config to only include content from the "
138
+ "domains used in the 'RealNews' dataset (Zellers et al., 2019).",
139
+ ),
140
+ C4Config(
141
+ language="en",
142
+ webtextlike=True,
143
+ description="Filters from the default config to only include content from the "
144
+ "URLs in OpenWebText (https://github.com/jcpeterson/openwebtext).",
145
+ ),
146
+ ]
147
+
148
+ def manual_download_instructions(self):
149
+ return """\
150
+ For the WebText-like config, you must manually download 'OpenWebText.zip'
151
+ (from https://mega.nz/#F!EZZD0YwJ!9_PlEQzdMVLaNdKv_ICNVQ) and the Common Crawl
152
+ WET files from August 2018 to July 2019
153
+ (https://commoncrawl.org/the-data/get-started/) and place them in the
154
+ `data_dir`.
155
+ """
156
+
157
+ def _info(self):
158
+ features = {
159
+ "text": datasets.Value("string"),
160
+ "url": datasets.Value("string"),
161
+ "content-type": datasets.Value("string"),
162
+ "content-length": datasets.Value("string"),
163
+ "timestamp": datasets.Value("string"),
164
+ }
165
+ return datasets.DatasetInfo(
166
+ description=_DESCRIPTION,
167
+ features=datasets.Features(features),
168
+ citation=_CITATION,
169
+ homepage="https://github.com/google-research/text-to-text-transfer-transformer#datasets",
170
+ )
171
+
172
+ def _split_generators(self, dl_manager, pipeline):
173
+ import apache_beam as beam
174
+
175
+ # We will automatically down the default CC version(s), but others need to
176
+ # be manually downloaded.
177
+ cc_versions = set(self.config.cc_versions)
178
+ auto_cc_versions = cc_versions & set(_DEFAULT_CC_VERSIONS)
179
+ manual_cc_versions = cc_versions - set(_DEFAULT_CC_VERSIONS)
180
+
181
+ files_to_download = {}
182
+ files_to_download["wet_path_urls"] = [
183
+ _WET_PATH_URL.format(cc_version=cc_version) for cc_version in auto_cc_versions
184
+ ]
185
+ if self.config.clean:
186
+ files_to_download["badwords"] = _BADWORDS_URL.format(lang=self.config.lang)
187
+ if self.config.realnewslike:
188
+ files_to_download["realnews_domains"] = _REALNEWS_DOMAINS_URL
189
+ file_paths = dl_manager.download_and_extract(files_to_download)
190
+
191
+ if self.config.webtextlike:
192
+ owt_path = os.path.join(dl_manager.manual_dir, _OPENWEBTEXT_URLS_ZIP)
193
+ if not os.path.exists(owt_path):
194
+ raise FileNotFoundError(
195
+ "{} does not exist. Make sure you insert a manual dir via `datasets.load_dataset('c4', data_dir=...)` that includes a file name {}. Manual download instructions: {})".format(
196
+ owt_path, _OPENWEBTEXT_URLS_ZIP, self.manual_download_instructions
197
+ )
198
+ )
199
+ file_paths["openwebtext_urls_zip"] = dl_manager.extract(owt_path)
200
+
201
+ wet_urls = []
202
+ for wet_path_url in file_paths["wet_path_urls"]:
203
+ with open(wet_path_url, "r", encoding="utf-8") as f:
204
+ wet_urls.extend(["%s/%s" % (_DOWNLOAD_HOST, l.strip()) for l in f])
205
+ file_paths["wet_urls"] = wet_urls
206
+ file_paths["wet_files"] = []
207
+
208
+ for cc_version in manual_cc_versions:
209
+ cc_dir = os.path.join(dl_manager.manual_dir, cc_version)
210
+ wet_files = beam.io.filesystems.FileSystems.match(os.path.join(cc_dir, "*.warc.wet.gz"))
211
+ if not os.path.exists(cc_dir):
212
+ raise FileNotFoundError(
213
+ "{} does not exist. Make sure you insert a manual dir via `datasets.load_dataset('c4', data_dir=...)` that includes the files {}. Manual download instructions: {})".format(
214
+ cc_dir, "*.warc.wet.gz", self.manual_download_instructions
215
+ )
216
+ )
217
+ logging.info("Adding %d WET files for manually downloaded version %s.", len(wet_files), cc_version)
218
+ file_paths["wet_files"].extend(wet_files)
219
+
220
+ page_content_pcollection = self._get_page_content(pipeline, file_paths, dl_manager)
221
+ return [
222
+ datasets.SplitGenerator(
223
+ name=datasets.Split.TRAIN,
224
+ gen_kwargs=dict(
225
+ split="train",
226
+ page_content=page_content_pcollection,
227
+ hashed_url_predicate=lambda x: x % 1000 != 0, # 99.9%
228
+ ),
229
+ ),
230
+ datasets.SplitGenerator(
231
+ name=datasets.Split.VALIDATION,
232
+ gen_kwargs=dict(
233
+ split="validation",
234
+ page_content=page_content_pcollection,
235
+ hashed_url_predicate=lambda x: x % 1000 == 0, # 0.01%
236
+ ),
237
+ ),
238
+ ]
239
+
240
+ def _get_page_content(self, pipeline, file_paths, dl_manager):
241
+ """Build PCollection of un-split page content."""
242
+ import apache_beam as beam
243
+
244
+ wet_file_paths = pipeline | "create_wet_files" >> beam.Create(file_paths["wet_files"])
245
+ if "wet_urls" in file_paths:
246
+
247
+ def download_url(url, downloader, pipeline):
248
+ path = downloader.download(url)
249
+ if not pipeline.is_local():
250
+ path = downloader.ship_files_with_pipeline(path, pipeline)
251
+ return path
252
+
253
+ dl_wet_file_paths = (
254
+ pipeline
255
+ | "create_wet_urls" >> beam.Create(file_paths["wet_urls"])
256
+ | beam.Map(download_url, downloader=dl_manager, pipeline=pipeline)
257
+ )
258
+ wet_file_paths = (wet_file_paths, dl_wet_file_paths) | beam.Flatten()
259
+
260
+ # Parse WET files and filter by length.
261
+ # Output: url, text
262
+ page_content = wet_file_paths | beam.FlatMap(split_wet_file) | beam.Filter(is_valid_length)
263
+
264
+ # Optionally filter for RealNews domains.
265
+ # Output: url, text
266
+ if self.config.realnewslike:
267
+ with open(file_paths["realnews_domains"], "r", encoding="utf-8") as f:
268
+ realnews_domains = json.load(f)
269
+ page_content = page_content | beam.Filter(is_realnews_domain, realnews_domains)
270
+
271
+ # Normalize and deduplicate by URL.
272
+ # Output: url, text
273
+ page_content = (
274
+ page_content
275
+ | "normalize_url" >> beam.Map(normalize_url)
276
+ | "group_url" >> beam.GroupByKey()
277
+ | beam.Map(dedupe_urls)
278
+ )
279
+
280
+ # Optionally filter for WebText-like URLs.
281
+ # Output: url, text
282
+ if self.config.webtextlike:
283
+ webtextlike_urls = (
284
+ pipeline
285
+ | "read_webtextlike_urls"
286
+ >> beam.io.ReadFromText(
287
+ os.path.join(file_paths["openwebtext_urls_zip"], _OPENWEBTEXT_URLS_FILE_PATTERN)
288
+ )
289
+ | "add_dummy_page" >> beam.Map(lambda x: (x, ""))
290
+ | "normal_webtext_url" >> beam.Map(normalize_url)
291
+ )
292
+ page_content = (
293
+ {"text": page_content, "webtextlike_urls": webtextlike_urls}
294
+ | "group_webtextlike_urls" >> beam.CoGroupByKey()
295
+ | beam.FlatMap(filter_by_webtextlike)
296
+ )
297
+
298
+ # Optionally clean pages of badwords, boilerpolate text, and duplicate
299
+ # spans of sentences.
300
+ # Output: url, text
301
+ if self.config.clean:
302
+ with open(file_paths["badwords"], "r", encoding="utf-8") as f:
303
+ badwords = [l.strip() for l in f]
304
+ page_content = page_content | "clean_pages" >> beam.FlatMap(get_clean_page_fn(badwords))
305
+ page_content = remove_duplicate_text(page_content)
306
+
307
+ # Optionally filter out non-`language` pages. We do this after cleaning
308
+ # since it may change the predominate language.
309
+ if self.config.lang != "all":
310
+ page_content |= beam.Filter(is_language, language=self.config.lang)
311
+
312
+ return page_content
313
+
314
+ def _build_pcollection(self, unused_pipeline, split, page_content, hashed_url_predicate):
315
+ import apache_beam as beam
316
+
317
+ def _emit_examples(el):
318
+ get_counter_inc_fn(split)("examples")
319
+ _, features = el
320
+ return (
321
+ features["url"],
322
+ {
323
+ "url": features["url"],
324
+ "text": features["text"],
325
+ "content-type": features["content-type"],
326
+ "content-length": features["content-length"],
327
+ "timestamp": features["timestamp"],
328
+ },
329
+ )
330
+
331
+ return page_content | beam.Filter(get_hashed_url_filter_fn(hashed_url_predicate)) | beam.Map(_emit_examples)
c4_utils.py ADDED
@@ -0,0 +1,489 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2020 The TensorFlow Datasets Authors and the HuggingFace Datasets Authors.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ # Lint as: python3
17
+ """Utilities for generating the C4 dataset."""
18
+
19
+ from __future__ import absolute_import, division, print_function
20
+
21
+ import functools
22
+ import gzip
23
+ import hashlib
24
+ import io
25
+ import re
26
+ import threading
27
+
28
+
29
+ # WET file constants
30
+ _PAGE_DELIMITER = "WARC/1.0"
31
+ _URL_KEY = "WARC-Target-URI:"
32
+ _URL_DATE = "WARC-Date:"
33
+ _CONTENT_TYPE = "Content-Type:"
34
+ _CONTENT_LEN = "Content-Length:"
35
+ _METADATA_PREFIXES = ("WARC", "CONTENT-", "Content-")
36
+
37
+ # Filters
38
+ _MIN_WORDS_PER_LINE = 5
39
+ _MIN_NUM_SENTENCES = 3
40
+ _MAX_WORD_LENGTH = 1000
41
+ _END_MARKS = (".", "?", "!", '"')
42
+ _ELLIPSIS = "..."
43
+ _POLICY_SUBSTRINGS = [
44
+ "terms of use",
45
+ "privacy policy",
46
+ "cookie policy",
47
+ "uses cookies",
48
+ "use of cookies",
49
+ "use cookies",
50
+ ]
51
+
52
+ # Memoized sentence tokenizer.
53
+ _SENTENCE_TOKENIZER = None
54
+
55
+
56
+ def get_counter_inc_fn(namespace):
57
+ import apache_beam as beam
58
+
59
+ def counter_inc_fn(counter, amt=1):
60
+ beam.metrics.Metrics.counter(namespace, counter).inc(amt)
61
+
62
+ return counter_inc_fn
63
+
64
+
65
+ def get_hashed_url_filter_fn(predicate_fn):
66
+ import tensorflow.compat.v2 as tf
67
+
68
+ def filter_fn(el):
69
+ url, _ = el
70
+ val = int(hashlib.md5(tf.compat.as_text(url).encode("utf-8")).hexdigest(), 16)
71
+ return predicate_fn(val)
72
+
73
+ return filter_fn
74
+
75
+
76
+ def _load_sentence_tokenizer():
77
+ """Returns a sentence tokenization function."""
78
+ # Lock to avoid a race-condition in the creation of the download directory.
79
+ with threading.Lock():
80
+ import nltk
81
+
82
+ nltk.download("punkt")
83
+ return nltk.data.load("nltk:tokenizers/punkt/english.pickle")
84
+
85
+
86
+ def _get_sentences(text):
87
+ import tensorflow.compat.v2 as tf
88
+
89
+ global _SENTENCE_TOKENIZER
90
+ if not _SENTENCE_TOKENIZER:
91
+ _SENTENCE_TOKENIZER = _load_sentence_tokenizer()
92
+ return list(_SENTENCE_TOKENIZER.tokenize(tf.compat.as_text(text)))
93
+
94
+
95
+ def _get_sentences_by_line(text, lower=False):
96
+ sentences = []
97
+ for line in text.splitlines():
98
+ sentences.append([s.lower() if lower else s for s in _get_sentences(line)])
99
+ return sentences
100
+
101
+
102
+ def is_language(page, language, min_probability=0.99):
103
+ """Returns True iff text is in `language` with at least `min_probability`."""
104
+ unused_url, features = page
105
+ text = features["text"]
106
+
107
+ counter_inc_fn = get_counter_inc_fn("detected-lang")
108
+
109
+ # Make langdetect predictions deterministic.
110
+ import langdetect
111
+
112
+ langdetect.DetectorFactory.seed = 0
113
+ try:
114
+ predictions = langdetect.detect_langs(text)
115
+ except langdetect.lang_detect_exception.LangDetectException:
116
+ counter_inc_fn("langdetect-exception")
117
+ return False
118
+ if not predictions:
119
+ counter_inc_fn("page-filtered-nolangpredictions")
120
+ return False
121
+ best_prediction = predictions[0]
122
+ if best_prediction.prob < min_probability:
123
+ counter_inc_fn("page-filtered-lowlangdetectconf")
124
+ return False
125
+ if best_prediction.lang != language:
126
+ counter_inc_fn("page-filtered-ignoredlang")
127
+ counter_inc_fn("page-filtered-ignoredlang-%s" % (best_prediction.lang))
128
+ return False
129
+ counter_inc_fn("page-emited-%s" % best_prediction.lang)
130
+ return True
131
+
132
+
133
+ def get_clean_page_fn(badwords=None):
134
+ """Returns `clean_page` with pre-compiled badword and citation regexes."""
135
+ # Used to filter citation from Wikipedia pages (among others).
136
+ citation_regex = re.compile(r"\[\d*\]|\[edit\]|\[citation needed\]")
137
+ if badwords:
138
+ badwords_regex = re.compile("[^a-z]({})[^a-z]".format("|".join(badwords or [])))
139
+ else:
140
+ badwords_regex = None
141
+ return functools.partial(clean_page, citation_regex=citation_regex, badwords_regex=badwords_regex)
142
+
143
+
144
+ def clean_page(
145
+ url_and_features,
146
+ citation_regex,
147
+ badwords_regex=None,
148
+ counter_inc_fn=None,
149
+ min_words_per_line=_MIN_WORDS_PER_LINE,
150
+ min_num_sentences=_MIN_NUM_SENTENCES,
151
+ max_word_length=_MAX_WORD_LENGTH,
152
+ ):
153
+ """Cleans a CommonCrawl page, yielding nothing if it should be skipped.
154
+
155
+ Cleaning removes lines with no end marks or with too few words. After line
156
+ filtering, pages are filtered out if they have too few sentences based on a
157
+ simple count of end marks.
158
+
159
+ Args:
160
+ url_and_features: tuple(string, dict), the url and features of the page.
161
+ citation_regex: Regex to use for finding Wikipedia-like citations to filter.
162
+ badwords_regex: Regex to use for finding badwords. Default None, which means
163
+ don't apply badwords filtering.
164
+ counter_inc_fn: function, a function taking the name of a counter to be
165
+ incremented and the (optional) amount. Defaults to a beam Metric counter.
166
+ min_words_per_line: int, the minimum number of words a line needs to not be
167
+ removed.
168
+ min_num_sentences: int, the minimum number of sentences a page needs to not
169
+ be skipped.
170
+ max_word_length: int, the maximum number of characters allowed in a word.
171
+ Lines containing a word with too many characters are removed.
172
+ Yields:
173
+ The url and cleaned text for the page.
174
+ """
175
+ url, features = url_and_features
176
+ text = features["text"]
177
+
178
+ if not counter_inc_fn:
179
+ counter_inc_fn = get_counter_inc_fn("clean-page")
180
+
181
+ lines = text.splitlines()
182
+ valid_lines = []
183
+ num_sentences = 0
184
+
185
+ def line_has_too_long_word(line):
186
+ for word in line.split():
187
+ if len(word) > max_word_length:
188
+ return True
189
+ return False
190
+
191
+ for line in lines:
192
+ line = line.strip()
193
+ if line_has_too_long_word(line):
194
+ counter_inc_fn("lines-with-too-long-word")
195
+ continue
196
+ line = citation_regex.sub("", line)
197
+ if not line.endswith(_END_MARKS) or line.endswith(_ELLIPSIS):
198
+ counter_inc_fn("lines-no-endmark")
199
+ continue
200
+ if len(line.split()) < min_words_per_line:
201
+ counter_inc_fn("lines-too-short")
202
+ continue
203
+ line_lower = line.lower()
204
+ # Remove documents which contain lorem ipsum
205
+ if "lorem ipsum" in line_lower:
206
+ counter_inc_fn("filtered-page-loremipsum")
207
+ return
208
+ # Remove "javascript must be enabled" notices
209
+ if "javascript" in line_lower:
210
+ counter_inc_fn("lines-javascript")
211
+ continue
212
+ # Remove docs which probably contain javascript code
213
+ if "{" in line:
214
+ counter_inc_fn("filtered-page-squigglybracket")
215
+ return
216
+ # Remove policy lines
217
+ if any(p in line_lower for p in _POLICY_SUBSTRINGS):
218
+ counter_inc_fn("lines-policy")
219
+ continue
220
+ # If any badword appears on its own in the line, skip this doc
221
+ if badwords_regex:
222
+ badwords_found = badwords_regex.search(line_lower)
223
+ if badwords_found is not None:
224
+ counter_inc_fn("filtered-page-badword")
225
+ return
226
+ num_sentences += len(_get_sentences(line))
227
+ valid_lines.append(line)
228
+ counter_inc_fn("lines-valid")
229
+
230
+ if num_sentences < min_num_sentences:
231
+ counter_inc_fn("filtered-page-toofewsentences")
232
+ return
233
+ counter_inc_fn("emitted-clean-pages")
234
+ features["text"] = "\n".join(valid_lines).strip()
235
+ yield url, features
236
+
237
+
238
+ def _hash_line(line):
239
+ import tensorflow.compat.v2 as tf
240
+
241
+ m = hashlib.md5()
242
+ m.update(tf.compat.as_text(line).encode("utf-8").strip().lower())
243
+ return m.hexdigest()
244
+
245
+
246
+ def _emit_url_to_lines(page):
247
+ """Emits url to all (lower-cased, hashed) lines."""
248
+ url, features = page
249
+ text = features["text"]
250
+ for line in text.split("\n"):
251
+ yield _hash_line(line), url
252
+
253
+
254
+ def _emit_line_to_urls(el, counter_inc_fn):
255
+ """Emits (hashed) line to all but one url."""
256
+ import tensorflow.compat.v2 as tf
257
+
258
+ line, urls = el
259
+ # Materialize urls as a list.
260
+ urls = list(urls)
261
+ # Hash urls and sort to have a consistent, but unbiased, selection when the
262
+ # same urls exist for multiple lines.
263
+ skip_url = min(urls, key=lambda x: hashlib.md5(tf.compat.as_text(x).encode("utf-8")).hexdigest())
264
+ for url in urls:
265
+ if url != skip_url:
266
+ yield url, line
267
+ counter_inc_fn("emitted-line-duplicate", amt=len(urls) - 1)
268
+
269
+
270
+ def _remove_lines_from_text(el, counter_inc_fn, min_num_sentences=_MIN_NUM_SENTENCES):
271
+ """Removes matching lines from the page.
272
+
273
+ Process the result of a join containing a single value for 'features' and zero
274
+ or more values for 'lines'. Each value in 'lines' is a lower-cased, hashed
275
+ line.
276
+
277
+ If a line has fewer sentences than `max_window_size`, the full line is
278
+ compared for a match.
279
+
280
+ Args:
281
+ el: `(string, {'features': features_dict, 'lines': [string]})`,
282
+ element containing the result of a join on key with both the page text
283
+ and lower-cased, hashed lines to remove.
284
+ counter_inc_fn: function, a function taking the name of a counter to be
285
+ incremented and the (optional) amount.
286
+ min_num_sentences: int, the minimum number of sentences a page needs to not
287
+ be skipped.
288
+
289
+ Yields:
290
+ url: The URL of the page.
291
+ features: The page features with lines removed from text.
292
+ """
293
+ url, join_values = el
294
+ features = join_values["features"]
295
+
296
+ assert len(features) == 1, "Invalid page count (%d) for %s" % (len(features), url)
297
+ features = features[0]
298
+ text = features["text"]
299
+ lines_to_remove = set(join_values["lines"])
300
+ new_lines = []
301
+ hashed_lines = set()
302
+ for line in text.split("\n"):
303
+ hashed_line = _hash_line(line)
304
+ if hashed_line in lines_to_remove:
305
+ counter_inc_fn("filtered-lines-duplicate")
306
+ elif hashed_line not in hashed_lines:
307
+ new_lines.append(line)
308
+ hashed_lines.add(hashed_line)
309
+ new_text = "\n".join(new_lines)
310
+ if len(_get_sentences(new_text)) < min_num_sentences:
311
+ counter_inc_fn("filtered-doc-toofewsentences")
312
+ return
313
+ new_features = features.copy()
314
+ new_features["text"] = new_text
315
+ yield (url, new_features)
316
+
317
+
318
+ def remove_duplicate_text(pages):
319
+ """Utility to remove duplicate lines across text documents."""
320
+ # Output: url, lines
321
+ import apache_beam as beam
322
+
323
+ counter_inc_fn = get_counter_inc_fn("dedupe-lines")
324
+ lines_to_remove = (
325
+ pages
326
+ | beam.FlatMap(_emit_url_to_lines)
327
+ | "group_sentences" >> beam.GroupByKey()
328
+ | beam.FlatMap(_emit_line_to_urls, counter_inc_fn=counter_inc_fn)
329
+ )
330
+
331
+ # Output: url, text
332
+ final_docs = (
333
+ {"features": pages, "lines": lines_to_remove}
334
+ | "group_features_and_lines_by_url" >> beam.CoGroupByKey()
335
+ | beam.FlatMap(_remove_lines_from_text, counter_inc_fn=counter_inc_fn)
336
+ )
337
+
338
+ return final_docs
339
+
340
+
341
+ def split_wet_file(wet_file_path, counter_inc_fn=None):
342
+ """Split a WET file into separate pages."""
343
+ from absl import logging
344
+
345
+ logging.info("Splitting file: %s", wet_file_path)
346
+ if not counter_inc_fn:
347
+ counter_inc_fn = get_counter_inc_fn("split-wet-file")
348
+ counter_inc_fn("wet-file")
349
+
350
+ import apache_beam as beam
351
+
352
+ with beam.io.filesystems.FileSystems.open(wet_file_path) as f, gzip.GzipFile(fileobj=f) as g:
353
+ url = None
354
+ content = None
355
+ content_len = None
356
+ content_type = None
357
+ timestamp = None
358
+
359
+ def _maybe_get_page():
360
+ """Generate a (url, {features}) page."""
361
+ if not url and url is not None:
362
+ counter_inc_fn("page-filtered-nourl")
363
+ if not content and content is not None:
364
+ counter_inc_fn("page-filtered-nocontent")
365
+ if not content_type and content_type is not None:
366
+ counter_inc_fn("page-nocontenttype")
367
+ if not content_len and content_len is not None:
368
+ counter_inc_fn("page-nocontentlen")
369
+ if not timestamp and timestamp is not None:
370
+ counter_inc_fn("page-notimestamp")
371
+ if content and url:
372
+ counter_inc_fn("page-emitted")
373
+ return (
374
+ url,
375
+ {
376
+ "text": "\n".join(content),
377
+ "content-type": content_type,
378
+ "content-length": content_len,
379
+ "timestamp": timestamp,
380
+ "url": url,
381
+ },
382
+ )
383
+ return None
384
+
385
+ for line in io.TextIOWrapper(g, encoding="utf-8"):
386
+ line = line.strip()
387
+ if not line:
388
+ continue
389
+ if line == _PAGE_DELIMITER:
390
+ page = _maybe_get_page()
391
+ if page:
392
+ yield page
393
+ url = ""
394
+ content = []
395
+ content_len = ""
396
+ content_type = ""
397
+ timestamp = ""
398
+
399
+ if line.startswith(_URL_KEY):
400
+ url = line[len(_URL_KEY) :].strip()
401
+
402
+ if line.startswith(_URL_DATE):
403
+ timestamp = line[len(_URL_DATE) :].strip()
404
+
405
+ if line.startswith(_CONTENT_TYPE):
406
+ content_type = line[len(_CONTENT_TYPE) :].strip()
407
+
408
+ if line.startswith(_CONTENT_LEN):
409
+ content_len = line[len(_CONTENT_LEN) :].strip()
410
+
411
+ if line.startswith(_METADATA_PREFIXES):
412
+ continue
413
+
414
+ content.append(line)
415
+
416
+ page = _maybe_get_page()
417
+ if page:
418
+ yield page
419
+
420
+
421
+ def dedupe_urls(el):
422
+ """Returns the first value for a given URL."""
423
+ counter_inc_fn = get_counter_inc_fn("dedupe-urls")
424
+ url, vals = el
425
+ cnt = 0
426
+ v = None
427
+ for v in vals:
428
+ cnt += 1
429
+ counter_inc_fn("filtered-url-duplicate", cnt - 1)
430
+ counter_inc_fn("unique-url")
431
+ return url, v
432
+
433
+
434
+ def is_valid_length(el, max_length=1.9e5):
435
+ """Returns False iff page's text is too long."""
436
+ counter_inc_fn = get_counter_inc_fn("is-valid-length")
437
+ _, page = el
438
+ if len(page["text"]) > max_length:
439
+ counter_inc_fn("filtered-page-contenttoolong")
440
+ return False
441
+ counter_inc_fn("valid-length")
442
+ return True
443
+
444
+
445
+ def is_realnews_domain(el, realnews_domains):
446
+ """Returns False iff page's (sub)domain is not allowed."""
447
+ import tldextract
448
+
449
+ counter_inc_fn = get_counter_inc_fn("is-realnews-domain")
450
+ url, _ = el
451
+ ext = tldextract.extract(url)
452
+ main_domain = ext.domain + "." + ext.suffix
453
+ if main_domain not in realnews_domains:
454
+ counter_inc_fn("filtered-url-invaliddomain")
455
+ return False
456
+ allowed_subdomains = realnews_domains[main_domain]
457
+ if isinstance(allowed_subdomains, list) and ext.subdomain not in allowed_subdomains:
458
+ counter_inc_fn("filtered-url-invalidsubdomain")
459
+ return False
460
+ counter_inc_fn("realnews-domain")
461
+ return True
462
+
463
+
464
+ def filter_by_webtextlike(el):
465
+ """Yields only pages with a matching WebText-like URL."""
466
+ counter_inc_fn = get_counter_inc_fn("filter-by-webtextlike")
467
+ url, join_values = el
468
+ text = join_values["text"]
469
+ webtextlike = join_values["webtextlike_urls"]
470
+ if not webtextlike:
471
+ counter_inc_fn("filtered-url-notwebtextlike")
472
+ return
473
+ if not text:
474
+ counter_inc_fn("missing-webtextlike")
475
+ return
476
+ assert len(text) == 1
477
+ counter_inc_fn("found-webtextlike")
478
+ yield url, text[0]
479
+
480
+
481
+ def normalize_url(el):
482
+ import tensorflow.compat.v2 as tf
483
+
484
+ url, val = el
485
+ url = tf.compat.as_text(url)
486
+ url = re.sub(r"https?:\/\/(www\.)?", "", url)
487
+ url = re.sub(r"\?(utm_|ref|feed).*", "", url)
488
+ url = url.rstrip("/")
489
+ return url, val