yuyang commited on
Commit
38fd326
1 Parent(s): 54e9da2

add data script and README

Browse files
Files changed (2) hide show
  1. README.md +7 -3
  2. bart_cnndm.py +275 -0
README.md CHANGED
@@ -1,3 +1,7 @@
1
- ---
2
- license: mit
3
- ---
 
 
 
 
 
1
+ Modification of the cnn_dailymail dataset in Hugging Face. The main goal is to reproduce the results on BART.
2
+
3
+ References: https://github.com/facebookresearch/fairseq/issues/1401
4
+
5
+ Major changes:
6
+ 1. remove the space in " ." in fix_missing_period.
7
+ 2. remove "(CNN)" in article.
bart_cnndm.py ADDED
@@ -0,0 +1,275 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+
18
+
19
+ """CNN/DailyMail Summarization dataset, non-anonymized version for BART model.
20
+
21
+ Major changes:
22
+ 1. remove the space in " ." in fix_missing_period.
23
+ 2. remove "(CNN)" in article.
24
+
25
+ The goal is to reproduce the results in BART.
26
+
27
+ References: https://github.com/facebookresearch/fairseq/issues/1401
28
+ """
29
+
30
+ import hashlib
31
+ import os
32
+
33
+ import datasets
34
+
35
+
36
+ logger = datasets.logging.get_logger(__name__)
37
+
38
+
39
+ _HOMEPAGE = "https://github.com/abisee/cnn-dailymail"
40
+
41
+ _DESCRIPTION = """\
42
+ CNN/DailyMail non-anonymized summarization dataset.
43
+ There are two features:
44
+ - article: text of news article, used as the document to be summarized
45
+ - highlights: joined text of highlights with <s> and </s> around each
46
+ highlight, which is the target summary
47
+ """
48
+
49
+ # The second citation introduces the source data, while the first
50
+ # introduces the specific form (non-anonymized) we use here.
51
+ _CITATION = """\
52
+ @article{DBLP:journals/corr/SeeLM17,
53
+ author = {Abigail See and
54
+ Peter J. Liu and
55
+ Christopher D. Manning},
56
+ title = {Get To The Point: Summarization with Pointer-Generator Networks},
57
+ journal = {CoRR},
58
+ volume = {abs/1704.04368},
59
+ year = {2017},
60
+ url = {http://arxiv.org/abs/1704.04368},
61
+ archivePrefix = {arXiv},
62
+ eprint = {1704.04368},
63
+ timestamp = {Mon, 13 Aug 2018 16:46:08 +0200},
64
+ biburl = {https://dblp.org/rec/bib/journals/corr/SeeLM17},
65
+ bibsource = {dblp computer science bibliography, https://dblp.org}
66
+ }
67
+ @inproceedings{hermann2015teaching,
68
+ title={Teaching machines to read and comprehend},
69
+ author={Hermann, Karl Moritz and Kocisky, Tomas and Grefenstette, Edward and Espeholt, Lasse and Kay, Will and Suleyman, Mustafa and Blunsom, Phil},
70
+ booktitle={Advances in neural information processing systems},
71
+ pages={1693--1701},
72
+ year={2015}
73
+ }
74
+ """
75
+
76
+ _DL_URLS = {
77
+ "cnn_stories": "https://huggingface.co/datasets/cnn_dailymail/resolve/11343c3752184397d56efc19a8a7cceb68089318/data/cnn_stories.tgz",
78
+ "dm_stories": "https://huggingface.co/datasets/cnn_dailymail/resolve/11343c3752184397d56efc19a8a7cceb68089318/data/dailymail_stories.tgz",
79
+ "train": "https://raw.githubusercontent.com/abisee/cnn-dailymail/master/url_lists/all_train.txt",
80
+ "validation": "https://raw.githubusercontent.com/abisee/cnn-dailymail/master/url_lists/all_val.txt",
81
+ "test": "https://raw.githubusercontent.com/abisee/cnn-dailymail/master/url_lists/all_test.txt",
82
+ }
83
+
84
+ _HIGHLIGHTS = "highlights"
85
+ _ARTICLE = "article"
86
+
87
+ _SUPPORTED_VERSIONS = [
88
+ # Using cased version.
89
+ datasets.Version("3.0.0", "Using cased version."),
90
+ # Same data as 0.0.2
91
+ datasets.Version("1.0.0", ""),
92
+ # Having the model predict newline separators makes it easier to evaluate
93
+ # using summary-level ROUGE.
94
+ datasets.Version("2.0.0", "Separate target sentences with newline."),
95
+ ]
96
+
97
+
98
+ _DEFAULT_VERSION = datasets.Version("3.0.0", "Using cased version.")
99
+
100
+
101
+ class CnnDailymailConfig(datasets.BuilderConfig):
102
+ """BuilderConfig for CnnDailymail."""
103
+
104
+ def __init__(self, **kwargs):
105
+ """BuilderConfig for CnnDailymail.
106
+ Args:
107
+ **kwargs: keyword arguments forwarded to super.
108
+ """
109
+ super(CnnDailymailConfig, self).__init__(**kwargs)
110
+
111
+
112
+ def _get_url_hashes(path):
113
+ """Get hashes of urls in file."""
114
+ urls = _read_text_file_path(path)
115
+
116
+ def url_hash(u):
117
+ h = hashlib.sha1()
118
+ try:
119
+ u = u.encode("utf-8")
120
+ except UnicodeDecodeError:
121
+ logger.error("Cannot hash url: %s", u)
122
+ h.update(u)
123
+ return h.hexdigest()
124
+
125
+ return {url_hash(u) for u in urls}
126
+
127
+
128
+ def _get_hash_from_path(p):
129
+ """Extract hash from path."""
130
+ return os.path.splitext(os.path.basename(p))[0]
131
+
132
+
133
+ DM_SINGLE_CLOSE_QUOTE = "\u2019" # unicode
134
+ DM_DOUBLE_CLOSE_QUOTE = "\u201d"
135
+ # acceptable ways to end a sentence
136
+ END_TOKENS = [
137
+ ".",
138
+ "!",
139
+ "?",
140
+ "...",
141
+ "'",
142
+ "`",
143
+ '"',
144
+ DM_SINGLE_CLOSE_QUOTE,
145
+ DM_DOUBLE_CLOSE_QUOTE,
146
+ ")",
147
+ ]
148
+
149
+
150
+ def _read_text_file_path(path):
151
+ with open(path, "r", encoding="utf-8") as f:
152
+ lines = [line.strip() for line in f]
153
+ return lines
154
+
155
+
156
+ def _read_text_file(file):
157
+ return [line.decode("utf-8").strip() for line in file]
158
+
159
+
160
+ def _get_art_abs(story_file, tfds_version):
161
+ """Get abstract (highlights) and article from a story file path."""
162
+ # Based on https://github.com/abisee/cnn-dailymail/blob/master/
163
+ # make_datafiles.py
164
+
165
+ lines = _read_text_file(story_file)
166
+
167
+ # The github code lowercase the text and we removed it in 3.0.0.
168
+
169
+ # Put periods on the ends of lines that are missing them
170
+ # (this is a problem in the dataset because many image captions don't end in
171
+ # periods; consequently they end up in the body of the article as run-on
172
+ # sentences)
173
+ def fix_missing_period(line):
174
+ """Adds a period to a line that is missing a period."""
175
+ if "@highlight" in line:
176
+ return line
177
+ if not line:
178
+ return line
179
+ if line[-1] in END_TOKENS:
180
+ return line
181
+ return line + "."
182
+
183
+ lines = [fix_missing_period(line) for line in lines]
184
+
185
+ # Separate out article and abstract sentences
186
+ article_lines = []
187
+ highlights = []
188
+ next_is_highlight = False
189
+ for line in lines:
190
+ if not line:
191
+ continue # empty line
192
+ elif line.startswith("@highlight"):
193
+ next_is_highlight = True
194
+ elif next_is_highlight:
195
+ highlights.append(line)
196
+ else:
197
+ article_lines.append(line)
198
+
199
+ # Make article into a single string
200
+ article = " ".join(article_lines)
201
+
202
+ if tfds_version >= "2.0.0":
203
+ abstract = "\n".join(highlights)
204
+ else:
205
+ abstract = " ".join(highlights)
206
+
207
+ if article[:5] == "(CNN)":
208
+ article = article[5:]
209
+
210
+ return article, abstract
211
+
212
+
213
+ class CnnDailymail(datasets.GeneratorBasedBuilder):
214
+ """CNN/DailyMail non-anonymized summarization dataset."""
215
+
216
+ BUILDER_CONFIGS = [
217
+ CnnDailymailConfig(name=str(version), description="Plain text", version=version)
218
+ for version in _SUPPORTED_VERSIONS
219
+ ]
220
+
221
+ def _info(self):
222
+ return datasets.DatasetInfo(
223
+ description=_DESCRIPTION,
224
+ features=datasets.Features(
225
+ {
226
+ _ARTICLE: datasets.Value("string"),
227
+ _HIGHLIGHTS: datasets.Value("string"),
228
+ "id": datasets.Value("string"),
229
+ }
230
+ ),
231
+ supervised_keys=None,
232
+ homepage=_HOMEPAGE,
233
+ citation=_CITATION,
234
+ )
235
+
236
+ def _vocab_text_gen(self, paths):
237
+ for _, ex in self._generate_examples(paths):
238
+ yield " ".join([ex[_ARTICLE], ex[_HIGHLIGHTS]])
239
+
240
+ def _split_generators(self, dl_manager):
241
+ dl_paths = dl_manager.download(_DL_URLS)
242
+ return [
243
+ datasets.SplitGenerator(
244
+ name=split,
245
+ gen_kwargs={
246
+ "urls_file": dl_paths[split],
247
+ "files_per_archive": [
248
+ dl_manager.iter_archive(dl_paths["cnn_stories"]),
249
+ dl_manager.iter_archive(dl_paths["dm_stories"]),
250
+ ],
251
+ },
252
+ )
253
+ for split in [
254
+ datasets.Split.TRAIN,
255
+ datasets.Split.VALIDATION,
256
+ datasets.Split.TEST,
257
+ ]
258
+ ]
259
+
260
+ def _generate_examples(self, urls_file, files_per_archive):
261
+ urls = _get_url_hashes(urls_file)
262
+ idx = 0
263
+ for files in files_per_archive:
264
+ for path, file in files:
265
+ hash_from_path = _get_hash_from_path(path)
266
+ if hash_from_path in urls:
267
+ article, highlights = _get_art_abs(file, self.config.version)
268
+ if not article or not highlights:
269
+ continue
270
+ yield idx, {
271
+ _ARTICLE: article,
272
+ _HIGHLIGHTS: highlights,
273
+ "id": hash_from_path,
274
+ }
275
+ idx += 1