Ian Porada commited on
Commit
f1eaa06
1 Parent(s): d092e58

Revert "Delete superglue_wsc.py with huggingface_hub"

Browse files

This reverts commit 71b794c3c9c29a59622ff2f88489586673146a36.

Files changed (1) hide show
  1. superglue_wsc.py +354 -0
superglue_wsc.py ADDED
@@ -0,0 +1,354 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ """The WSC from the SuperGLUE benchmark."""
18
+
19
+
20
+ import json
21
+ import os
22
+
23
+ import datasets
24
+
25
+
26
+ _SUPER_GLUE_CITATION = """\
27
+ @article{wang2019superglue,
28
+ title={SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems},
29
+ author={Wang, Alex and Pruksachatkun, Yada and Nangia, Nikita and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R},
30
+ journal={arXiv preprint arXiv:1905.00537},
31
+ year={2019}
32
+ }
33
+
34
+ Note that each SuperGLUE dataset has its own citation. Please see the source to
35
+ get the correct citation for each contained dataset.
36
+ """
37
+
38
+ _GLUE_DESCRIPTION = """\
39
+ SuperGLUE (https://super.gluebenchmark.com/) is a new benchmark styled after
40
+ GLUE with a new set of more difficult language understanding tasks, improved
41
+ resources, and a new public leaderboard.
42
+
43
+ """
44
+
45
+ _WSC_DESCRIPTION = """\
46
+ The Winograd Schema Challenge (WSC, Levesque et al., 2012) is a reading comprehension
47
+ task in which a system must read a sentence with a pronoun and select the referent of that pronoun
48
+ from a list of choices. Given the difficulty of this task and the headroom still left, we have included
49
+ WSC in SuperGLUE and recast the dataset into its coreference form. The task is cast as a binary
50
+ classification problem, as opposed to N-multiple choice, in order to isolate the model's ability to
51
+ understand the coreference links within a sentence as opposed to various other strategies that may
52
+ come into play in multiple choice conditions. With that in mind, we create a split with 65% negative
53
+ majority class in the validation set, reflecting the distribution of the hidden test set, and 52% negative
54
+ class in the training set. The training and validation examples are drawn from the original Winograd
55
+ Schema dataset (Levesque et al., 2012), as well as those distributed by the affiliated organization
56
+ Commonsense Reasoning. The test examples are derived from fiction books and have been shared
57
+ with us by the authors of the original dataset. Previously, a version of WSC recast as NLI as included
58
+ in GLUE, known as WNLI. No substantial progress was made on WNLI, with many submissions
59
+ opting to submit only majority class predictions. WNLI was made especially difficult due to an
60
+ adversarial train/dev split: Premise sentences that appeared in the training set sometimes appeared
61
+ in the development set with a different hypothesis and a flipped label. If a system memorized the
62
+ training set without meaningfully generalizing, which was easy due to the small size of the training
63
+ set, it could perform far below chance on the development set. We remove this adversarial design
64
+ in the SuperGLUE version of WSC by ensuring that no sentences are shared between the training,
65
+ validation, and test sets.
66
+
67
+ However, the validation and test sets come from different domains, with the validation set consisting
68
+ of ambiguous examples such that changing one non-noun phrase word will change the coreference
69
+ dependencies in the sentence. The test set consists only of more straightforward examples, with a
70
+ high number of noun phrases (and thus more choices for the model), but low to no ambiguity."""
71
+
72
+
73
+ _WSC_CITATION = """\
74
+ @inproceedings{levesque2012winograd,
75
+ title={The winograd schema challenge},
76
+ author={Levesque, Hector and Davis, Ernest and Morgenstern, Leora},
77
+ booktitle={Thirteenth International Conference on the Principles of Knowledge Representation and Reasoning},
78
+ year={2012}
79
+ }"""
80
+
81
+
82
+ class SuperGlueConfig(datasets.BuilderConfig):
83
+ """BuilderConfig for SuperGLUE."""
84
+
85
+ def __init__(self, features, data_url, citation, url, label_classes=("False", "True"), **kwargs):
86
+ """BuilderConfig for SuperGLUE.
87
+
88
+ Args:
89
+ features: `list[string]`, list of the features that will appear in the
90
+ feature dict. Should not include "label".
91
+ data_url: `string`, url to download the zip file from.
92
+ citation: `string`, citation for the data set.
93
+ url: `string`, url for information about the data set.
94
+ label_classes: `list[string]`, the list of classes for the label if the
95
+ label is present as a string. Non-string labels will be cast to either
96
+ 'False' or 'True'.
97
+ **kwargs: keyword arguments forwarded to super.
98
+ """
99
+ # Version history:
100
+ # 1.0.3: Fix not including entity position in ReCoRD.
101
+ # 1.0.2: Fixed non-nondeterminism in ReCoRD.
102
+ # 1.0.1: Change from the pre-release trial version of SuperGLUE (v1.9) to
103
+ # the full release (v2.0).
104
+ # 1.0.0: S3 (new shuffling, sharding and slicing mechanism).
105
+ # 0.0.2: Initial version.
106
+ super(SuperGlueConfig, self).__init__(version=datasets.Version("1.0.3"), **kwargs)
107
+ self.features = features
108
+ self.label_classes = label_classes
109
+ self.data_url = data_url
110
+ self.citation = citation
111
+ self.url = url
112
+
113
+
114
+ class SuperGlue(datasets.GeneratorBasedBuilder):
115
+ """The SuperGLUE benchmark."""
116
+
117
+ BUILDER_CONFIGS = [
118
+ SuperGlueConfig(
119
+ name="wsc",
120
+ description=_WSC_DESCRIPTION,
121
+ # Note that span1_index and span2_index will be integers stored as
122
+ # datasets.Value('int32').
123
+ features=["text", "span1_index", "span2_index", "span1_text", "span2_text"],
124
+ data_url="https://dl.fbaipublicfiles.com/glue/superglue/data/v2/WSC.zip",
125
+ citation=_WSC_CITATION,
126
+ url="https://cs.nyu.edu/faculty/davise/papers/WinogradSchemas/WS.html",
127
+ ),
128
+ SuperGlueConfig(
129
+ name="wsc.fixed",
130
+ description=(
131
+ _WSC_DESCRIPTION + "\n\nThis version fixes issues where the spans are not actually "
132
+ "substrings of the text."
133
+ ),
134
+ # Note that span1_index and span2_index will be integers stored as
135
+ # datasets.Value('int32').
136
+ features=["text", "span1_index", "span2_index", "span1_text", "span2_text"],
137
+ data_url="https://dl.fbaipublicfiles.com/glue/superglue/data/v2/WSC.zip",
138
+ citation=_WSC_CITATION,
139
+ url="https://cs.nyu.edu/faculty/davise/papers/WinogradSchemas/WS.html",
140
+ ),
141
+ ]
142
+
143
+ def _info(self):
144
+ features = {feature: datasets.Value("string") for feature in self.config.features}
145
+ if self.config.name.startswith("wsc"):
146
+ features["span1_index"] = datasets.Value("int32")
147
+ features["span2_index"] = datasets.Value("int32")
148
+ if self.config.name == "wic":
149
+ features["start1"] = datasets.Value("int32")
150
+ features["start2"] = datasets.Value("int32")
151
+ features["end1"] = datasets.Value("int32")
152
+ features["end2"] = datasets.Value("int32")
153
+ if self.config.name == "multirc":
154
+ features["idx"] = dict(
155
+ {
156
+ "paragraph": datasets.Value("int32"),
157
+ "question": datasets.Value("int32"),
158
+ "answer": datasets.Value("int32"),
159
+ }
160
+ )
161
+ elif self.config.name == "record":
162
+ features["idx"] = dict(
163
+ {
164
+ "passage": datasets.Value("int32"),
165
+ "query": datasets.Value("int32"),
166
+ }
167
+ )
168
+ else:
169
+ features["idx"] = datasets.Value("int32")
170
+
171
+ if self.config.name == "record":
172
+ # Entities are the set of possible choices for the placeholder.
173
+ features["entities"] = datasets.features.Sequence(datasets.Value("string"))
174
+ # The start and end indices of paragraph text for each entity.
175
+ features["entity_spans"] = datasets.features.Sequence(
176
+ {
177
+ "text": datasets.Value("string"),
178
+ "start": datasets.Value("int32"),
179
+ "end": datasets.Value("int32"),
180
+ }
181
+ )
182
+ # Answers are the subset of entities that are correct.
183
+ features["answers"] = datasets.features.Sequence(datasets.Value("string"))
184
+ else:
185
+ features["label"] = datasets.features.ClassLabel(names=self.config.label_classes)
186
+
187
+ return datasets.DatasetInfo(
188
+ description=_GLUE_DESCRIPTION + self.config.description,
189
+ features=datasets.Features(features),
190
+ homepage=self.config.url,
191
+ citation=self.config.citation + "\n" + _SUPER_GLUE_CITATION,
192
+ )
193
+
194
+ def _split_generators(self, dl_manager):
195
+ dl_dir = dl_manager.download_and_extract(self.config.data_url) or ""
196
+ task_name = _get_task_name_from_data_url(self.config.data_url)
197
+ dl_dir = os.path.join(dl_dir, task_name)
198
+ if self.config.name in ["axb", "axg"]:
199
+ return [
200
+ datasets.SplitGenerator(
201
+ name=datasets.Split.TEST,
202
+ gen_kwargs={
203
+ "data_file": os.path.join(dl_dir, f"{task_name}.jsonl"),
204
+ "split": datasets.Split.TEST,
205
+ },
206
+ ),
207
+ ]
208
+ return [
209
+ datasets.SplitGenerator(
210
+ name=datasets.Split.TRAIN,
211
+ gen_kwargs={
212
+ "data_file": os.path.join(dl_dir, "train.jsonl"),
213
+ "split": datasets.Split.TRAIN,
214
+ },
215
+ ),
216
+ datasets.SplitGenerator(
217
+ name=datasets.Split.VALIDATION,
218
+ gen_kwargs={
219
+ "data_file": os.path.join(dl_dir, "val.jsonl"),
220
+ "split": datasets.Split.VALIDATION,
221
+ },
222
+ ),
223
+ datasets.SplitGenerator(
224
+ name=datasets.Split.TEST,
225
+ gen_kwargs={
226
+ "data_file": os.path.join(dl_dir, "test.jsonl"),
227
+ "split": datasets.Split.TEST,
228
+ },
229
+ ),
230
+ ]
231
+
232
+ def _generate_examples(self, data_file, split):
233
+ with open(data_file, encoding="utf-8") as f:
234
+ for line in f:
235
+ row = json.loads(line)
236
+
237
+ if self.config.name == "multirc":
238
+ paragraph = row["passage"]
239
+ for question in paragraph["questions"]:
240
+ for answer in question["answers"]:
241
+ label = answer.get("label")
242
+ key = "%s_%s_%s" % (row["idx"], question["idx"], answer["idx"])
243
+ yield key, {
244
+ "paragraph": paragraph["text"],
245
+ "question": question["question"],
246
+ "answer": answer["text"],
247
+ "label": -1 if label is None else _cast_label(bool(label)),
248
+ "idx": {"paragraph": row["idx"], "question": question["idx"], "answer": answer["idx"]},
249
+ }
250
+ elif self.config.name == "record":
251
+ passage = row["passage"]
252
+ entity_texts, entity_spans = _get_record_entities(passage)
253
+ for qa in row["qas"]:
254
+ yield qa["idx"], {
255
+ "passage": passage["text"],
256
+ "query": qa["query"],
257
+ "entities": entity_texts,
258
+ "entity_spans": entity_spans,
259
+ "answers": _get_record_answers(qa),
260
+ "idx": {"passage": row["idx"], "query": qa["idx"]},
261
+ }
262
+ else:
263
+ if self.config.name.startswith("wsc"):
264
+ row.update(row["target"])
265
+ example = {feature: row[feature] for feature in self.config.features}
266
+ if self.config.name == "wsc.fixed":
267
+ example = _fix_wst(example)
268
+ example["idx"] = row["idx"]
269
+
270
+ if "label" in row:
271
+ if self.config.name == "copa":
272
+ example["label"] = "choice2" if row["label"] else "choice1"
273
+ else:
274
+ example["label"] = _cast_label(row["label"])
275
+ else:
276
+ assert split == datasets.Split.TEST, row
277
+ example["label"] = -1
278
+ yield example["idx"], example
279
+
280
+
281
+ def _fix_wst(ex):
282
+ """Fixes most cases where spans are not actually substrings of text."""
283
+
284
+ def _fix_span_text(k):
285
+ """Fixes a single span."""
286
+ text = ex[k + "_text"]
287
+ index = ex[k + "_index"]
288
+
289
+ if text in ex["text"]:
290
+ return
291
+
292
+ if text in ("Kamenev and Zinoviev", "Kamenev, Zinoviev, and Stalin"):
293
+ # There is no way to correct these examples since the subjects have
294
+ # intervening text.
295
+ return
296
+
297
+ if "theyscold" in text:
298
+ ex["text"].replace("theyscold", "they scold")
299
+ ex["span2_index"] = 10
300
+ # Make sure case of the first words match.
301
+ first_word = ex["text"].split()[index]
302
+ if first_word[0].islower():
303
+ text = text[0].lower() + text[1:]
304
+ else:
305
+ text = text[0].upper() + text[1:]
306
+ # Remove punctuation in span.
307
+ text = text.rstrip(".")
308
+ # Replace incorrect whitespace character in span.
309
+ text = text.replace("\n", " ")
310
+ ex[k + "_text"] = text
311
+ assert ex[k + "_text"] in ex["text"], ex
312
+
313
+ _fix_span_text("span1")
314
+ _fix_span_text("span2")
315
+ return ex
316
+
317
+
318
+ def _cast_label(label):
319
+ """Converts the label into the appropriate string version."""
320
+ if isinstance(label, str):
321
+ return label
322
+ elif isinstance(label, bool):
323
+ return "True" if label else "False"
324
+ elif isinstance(label, int):
325
+ assert label in (0, 1)
326
+ return str(label)
327
+ else:
328
+ raise ValueError("Invalid label format.")
329
+
330
+
331
+ def _get_record_entities(passage):
332
+ """Returns the unique set of entities."""
333
+ text = passage["text"]
334
+ entity_spans = list()
335
+ for entity in passage["entities"]:
336
+ entity_text = text[entity["start"] : entity["end"] + 1]
337
+ entity_spans.append({"text": entity_text, "start": entity["start"], "end": entity["end"] + 1})
338
+ entity_spans = sorted(entity_spans, key=lambda e: e["start"]) # sort by start index
339
+ entity_texts = set(e["text"] for e in entity_spans) # for backward compatability
340
+ return entity_texts, entity_spans
341
+
342
+
343
+ def _get_record_answers(qa):
344
+ """Returns the unique set of answers."""
345
+ if "answers" not in qa:
346
+ return []
347
+ answers = set()
348
+ for answer in qa["answers"]:
349
+ answers.add(answer["text"])
350
+ return sorted(answers)
351
+
352
+
353
+ def _get_task_name_from_data_url(data_url):
354
+ return data_url.split("/")[-1].split(".")[0]