dfki-nlp commited on
Commit
a104c6e
1 Parent(s): fc96438

Upload multilingual_tacred.py

Browse files
Files changed (1) hide show
  1. multilingual_tacred.py +334 -0
multilingual_tacred.py ADDED
@@ -0,0 +1,334 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The current dataset script contributor.
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
+ """The TACRED Relation Classification dataset in various languages, DFKI format."""
17
+ import itertools
18
+ import json
19
+ import os
20
+
21
+ import datasets
22
+
23
+ _CITATION = """\
24
+ @inproceedings{zhang-etal-2017-position,
25
+ title = "Position-aware Attention and Supervised Data Improve Slot Filling",
26
+ author = "Zhang, Yuhao and
27
+ Zhong, Victor and
28
+ Chen, Danqi and
29
+ Angeli, Gabor and
30
+ Manning, Christopher D.",
31
+ booktitle = "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
32
+ month = sep,
33
+ year = "2017",
34
+ address = "Copenhagen, Denmark",
35
+ publisher = "Association for Computational Linguistics",
36
+ url = "https://www.aclweb.org/anthology/D17-1004",
37
+ doi = "10.18653/v1/D17-1004",
38
+ pages = "35--45",
39
+ }
40
+
41
+ @inproceedings{alt-etal-2020-tacred,
42
+ title = "{TACRED} Revisited: A Thorough Evaluation of the {TACRED} Relation Extraction Task",
43
+ author = "Alt, Christoph and
44
+ Gabryszak, Aleksandra and
45
+ Hennig, Leonhard",
46
+ booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
47
+ month = jul,
48
+ year = "2020",
49
+ address = "Online",
50
+ publisher = "Association for Computational Linguistics",
51
+ url = "https://www.aclweb.org/anthology/2020.acl-main.142",
52
+ doi = "10.18653/v1/2020.acl-main.142",
53
+ pages = "1558--1569",
54
+ }
55
+ """
56
+
57
+ _DESCRIPTION = """\
58
+ TACRED is a large-scale relation extraction dataset with 106,264 examples built over newswire
59
+ and web text from the corpus used in the yearly TAC Knowledge Base Population (TAC KBP) challenges.
60
+ Examples in TACRED cover 41 relation types as used in the TAC KBP challenges (e.g., per:schools_attended
61
+ and org:members) or are labeled as no_relation if no defined relation is held. These examples are created
62
+ by combining available human annotations from the TAC KBP challenges and crowdsourcing.
63
+
64
+ Please see our EMNLP paper, or our EMNLP slides for full details.
65
+
66
+ Note: There is currently a label-corrected version of the TACRED dataset, which you should consider using instead of
67
+ the original version released in 2017. For more details on this new version, see the TACRED Revisited paper
68
+ published at ACL 2020.
69
+
70
+ NOTE: This Datasetreader supports a reduced version of the original TACRED JSON format with the following changes:
71
+ - Removed fields: stanford_pos, stanford_ner, stanford_head, stanford_deprel, docid
72
+ The motivation for this is that we want to support additional languages, for which these fields were not required
73
+ or available. The reader expects the specification of a language-specific configuration specifying the variant
74
+ (original or revised) and the language (as a two-letter iso code). The default config is 'original-en'.
75
+
76
+ The Datasetreader changes the offsets of the following fields, to conform with standard Python usage (see
77
+ #_generate_examples()):
78
+ - subj_end to subj_end + 1 (make end offset exclusive)
79
+ - obj_end to obj_end + 1 (make end offset exclusive)
80
+ """
81
+
82
+ _HOMEPAGE = "https://nlp.stanford.edu/projects/tacred/"
83
+
84
+ _LICENSE = "LDC"
85
+
86
+ _URL = "https://catalog.ldc.upenn.edu/LDC2018T24"
87
+
88
+ # The HuggingFace dataset library don't host the datasets but only point to the original files
89
+ # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
90
+ _PATCH_URLs = {
91
+ "dev": "https://raw.githubusercontent.com/DFKI-NLP/tacrev/master/patch/dev_patch.json",
92
+ "test": "https://raw.githubusercontent.com/DFKI-NLP/tacrev/master/patch/test_patch.json",
93
+ }
94
+
95
+ _VERSION = datasets.Version("1.0.0")
96
+
97
+ _LANGS = [
98
+ "ar",
99
+ "de",
100
+ "en",
101
+ "es",
102
+ # "eu",
103
+ "fi",
104
+ "fr",
105
+ "hi",
106
+ "hu",
107
+ "ja",
108
+ "pl",
109
+ "ru",
110
+ "tr",
111
+ "zh",
112
+ ]
113
+
114
+ _CLASS_LABELS = [
115
+ "no_relation",
116
+ "org:alternate_names",
117
+ "org:city_of_headquarters",
118
+ "org:country_of_headquarters",
119
+ "org:dissolved",
120
+ "org:founded",
121
+ "org:founded_by",
122
+ "org:member_of",
123
+ "org:members",
124
+ "org:number_of_employees/members",
125
+ "org:parents",
126
+ "org:political/religious_affiliation",
127
+ "org:shareholders",
128
+ "org:stateorprovince_of_headquarters",
129
+ "org:subsidiaries",
130
+ "org:top_members/employees",
131
+ "org:website",
132
+ "per:age",
133
+ "per:alternate_names",
134
+ "per:cause_of_death",
135
+ "per:charges",
136
+ "per:children",
137
+ "per:cities_of_residence",
138
+ "per:city_of_birth",
139
+ "per:city_of_death",
140
+ "per:countries_of_residence",
141
+ "per:country_of_birth",
142
+ "per:country_of_death",
143
+ "per:date_of_birth",
144
+ "per:date_of_death",
145
+ "per:employee_of",
146
+ "per:origin",
147
+ "per:other_family",
148
+ "per:parents",
149
+ "per:religion",
150
+ "per:schools_attended",
151
+ "per:siblings",
152
+ "per:spouse",
153
+ "per:stateorprovince_of_birth",
154
+ "per:stateorprovince_of_death",
155
+ "per:stateorprovinces_of_residence",
156
+ "per:title",
157
+ ]
158
+
159
+
160
+ _NER_CLASS_LABELS = [
161
+ "LOCATION",
162
+ "ORGANIZATION",
163
+ "PERSON",
164
+ "DATE",
165
+ "MONEY",
166
+ "PERCENT",
167
+ "TIME",
168
+ "CAUSE_OF_DEATH",
169
+ "CITY",
170
+ "COUNTRY",
171
+ "CRIMINAL_CHARGE",
172
+ "EMAIL",
173
+ "HANDLE",
174
+ "IDEOLOGY",
175
+ "NATIONALITY",
176
+ "RELIGION",
177
+ "STATE_OR_PROVINCE",
178
+ "TITLE",
179
+ "URL",
180
+ "NUMBER",
181
+ "ORDINAL",
182
+ "MISC",
183
+ "DURATION",
184
+ "O",
185
+ ]
186
+
187
+
188
+ def convert_ptb_token(token: str) -> str:
189
+ """Convert PTB tokens to normal tokens"""
190
+ return {
191
+ "-lrb-": "(",
192
+ "-rrb-": ")",
193
+ "-lsb-": "[",
194
+ "-rsb-": "]",
195
+ "-lcb-": "{",
196
+ "-rcb-": "}",
197
+ }.get(token.lower(), token)
198
+
199
+
200
+ class TacredDfkiConfig(datasets.BuilderConfig):
201
+ def __init__(self, **kwargs):
202
+ super(TacredDfkiConfig, self).__init__(version=_VERSION, **kwargs)
203
+
204
+
205
+ class TacredDfki(datasets.GeneratorBasedBuilder):
206
+ """TACRED is a large-scale relation extraction dataset with 106,264 examples built over newswire
207
+ and web text from the corpus used in the yearly TAC Knowledge Base Population (TAC KBP) challenges."""
208
+
209
+ BUILDER_CONFIGS = [
210
+ TacredDfkiConfig(
211
+ name=f"{variant}-{lang}",
212
+ description=f"{'The revised TACRED (corrected labels in dev and test split)' if variant == 'revised' else 'The original TACRED'} examples in language '{lang}'.",
213
+ )
214
+ for (lang, variant) in itertools.product(_LANGS, ["original", "revised"])
215
+ ]
216
+
217
+ DEFAULT_CONFIG_NAME = "original-en" # type: ignore
218
+
219
+ @property
220
+ def manual_download_instructions(self):
221
+ return (
222
+ "To use TACRED you have to download it manually. "
223
+ "It is available via the LDC at https://catalog.ldc.upenn.edu/LDC2018T24"
224
+ "Please extract all files in one folder and load the dataset with: "
225
+ "`datasets.load_dataset('tacred', data_dir='path/to/folder/folder_name')`."
226
+ "Language-specific versions must be requested from git.nlp@dfki.de."
227
+ )
228
+
229
+ def _info(self):
230
+ features = datasets.Features(
231
+ {
232
+ "id": datasets.Value("string"),
233
+ "token": datasets.Sequence(datasets.Value("string")),
234
+ "subj_start": datasets.Value("int32"),
235
+ "subj_end": datasets.Value("int32"),
236
+ "subj_type": datasets.ClassLabel(names=_NER_CLASS_LABELS),
237
+ "obj_start": datasets.Value("int32"),
238
+ "obj_end": datasets.Value("int32"),
239
+ "obj_type": datasets.ClassLabel(names=_NER_CLASS_LABELS),
240
+ "relation": datasets.ClassLabel(names=_CLASS_LABELS),
241
+ }
242
+ )
243
+
244
+ return datasets.DatasetInfo(
245
+ # This is the description that will appear on the datasets page.
246
+ description=_DESCRIPTION,
247
+ # This defines the different columns of the dataset and their types
248
+ features=features, # Here we define them above because they are different between the two configurations
249
+ # If there's a common (input, target) tuple from the features,
250
+ # specify them here. They'll be used if as_supervised=True in
251
+ # builder.as_dataset.
252
+ supervised_keys=None,
253
+ # Homepage of the dataset for documentation
254
+ homepage=_HOMEPAGE,
255
+ # License for the dataset if available
256
+ license=_LICENSE,
257
+ # Citation for the dataset
258
+ citation=_CITATION,
259
+ )
260
+
261
+ def _split_generators(self, dl_manager):
262
+ """Returns SplitGenerators."""
263
+ # If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name
264
+
265
+ # dl_manager is a datasets.download.DownloadManager that can be used to download and extract URLs
266
+ # It can accept any type or nested list/dict and will give back the same structure with the url replaced with path to local files.
267
+ # By default the archives will be extracted and a path to a cached folder where they are extracted is returned instead of the archive
268
+ patch_files = {}
269
+ variant, lang = self.config.name.split("-")
270
+ if variant == "revised":
271
+ patch_files = dl_manager.download_and_extract(_PATCH_URLs)
272
+
273
+ data_dir = os.path.abspath(os.path.expanduser(dl_manager.manual_dir))
274
+
275
+ if not os.path.exists(data_dir):
276
+ raise FileNotFoundError(
277
+ "{} does not exist. Make sure you insert a manual dir via `datasets.load_dataset('DFKI-SLT/tacred_dfki', data_dir=...)` that includes the unzipped files from the TACRED_LDC zip. Manual download instructions: {}".format(
278
+ data_dir, self.manual_download_instructions
279
+ )
280
+ )
281
+
282
+ return [
283
+ datasets.SplitGenerator(
284
+ name=datasets.Split.TRAIN,
285
+ gen_kwargs={
286
+ "filepath": os.path.join(data_dir, lang, "train.json"),
287
+ "patch_filepath": None,
288
+ },
289
+ ),
290
+ datasets.SplitGenerator(
291
+ name=datasets.Split.TEST,
292
+ gen_kwargs={
293
+ "filepath": os.path.join(data_dir, lang, "test.json"),
294
+ "patch_filepath": patch_files.get("test"),
295
+ },
296
+ ),
297
+ datasets.SplitGenerator(
298
+ name=datasets.Split.VALIDATION,
299
+ gen_kwargs={
300
+ "filepath": os.path.join(data_dir, lang, "dev.json"),
301
+ "patch_filepath": patch_files.get("dev"),
302
+ },
303
+ ),
304
+ ]
305
+
306
+ def _generate_examples(self, filepath, patch_filepath):
307
+ """Yields examples."""
308
+ # This method will receive as arguments the `gen_kwargs` defined in the previous `_split_generators` method.
309
+ # It is in charge of opening the given file and yielding (key, example) tuples from the dataset
310
+ # The key is not important, it's more here for legacy reason (legacy from tfds)
311
+ patch_examples = {}
312
+ if patch_filepath is not None:
313
+ with open(patch_filepath, encoding="utf-8") as f:
314
+ patch_examples = {example["id"]: example for example in json.load(f)}
315
+
316
+ with open(filepath, encoding="utf-8") as f:
317
+ data = json.load(f)
318
+ for example in data:
319
+ id_ = example["id"]
320
+
321
+ if id_ in patch_examples:
322
+ example.update(patch_examples[id_])
323
+
324
+ yield id_, {
325
+ "id": example["id"],
326
+ "token": [convert_ptb_token(token) for token in example["token"]],
327
+ "subj_start": example["subj_start"],
328
+ "subj_end": example["subj_end"] + 1, # make end offset exclusive
329
+ "subj_type": example["subj_type"],
330
+ "obj_start": example["obj_start"],
331
+ "obj_end": example["obj_end"] + 1, # make end offset exclusive
332
+ "obj_type": example["obj_type"],
333
+ "relation": example["relation"],
334
+ }