Commit
•
ca822b4
1
Parent(s):
2adbb30
Delete loading script
Browse files- adv_glue.py +0 -330
adv_glue.py
DELETED
@@ -1,330 +0,0 @@
|
|
1 |
-
"""The Adversarial GLUE (AdvGLUE) benchmark.
|
2 |
-
Homepage: https://adversarialglue.github.io/
|
3 |
-
"""
|
4 |
-
import json
|
5 |
-
import os
|
6 |
-
import textwrap
|
7 |
-
|
8 |
-
import datasets
|
9 |
-
|
10 |
-
|
11 |
-
_ADV_GLUE_CITATION = """\
|
12 |
-
@article{Wang2021AdversarialGA,
|
13 |
-
title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},
|
14 |
-
author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},
|
15 |
-
journal={ArXiv},
|
16 |
-
year={2021},
|
17 |
-
volume={abs/2111.02840}
|
18 |
-
}
|
19 |
-
"""
|
20 |
-
|
21 |
-
_ADV_GLUE_DESCRIPTION = """\
|
22 |
-
Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark
|
23 |
-
that focuses on the adversarial robustness evaluation of language models. It covers five
|
24 |
-
natural language understanding tasks from the famous GLUE tasks and is an adversarial
|
25 |
-
version of GLUE benchmark.
|
26 |
-
"""
|
27 |
-
|
28 |
-
_MNLI_BASE_KWARGS = dict(
|
29 |
-
text_features={
|
30 |
-
"premise": "premise",
|
31 |
-
"hypothesis": "hypothesis",
|
32 |
-
},
|
33 |
-
label_classes=["entailment", "neutral", "contradiction"],
|
34 |
-
label_column="label",
|
35 |
-
data_url="https://dl.fbaipublicfiles.com/glue/data/MNLI.zip",
|
36 |
-
data_dir="MNLI",
|
37 |
-
citation=textwrap.dedent(
|
38 |
-
"""\
|
39 |
-
@InProceedings{N18-1101,
|
40 |
-
author = "Williams, Adina
|
41 |
-
and Nangia, Nikita
|
42 |
-
and Bowman, Samuel",
|
43 |
-
title = "A Broad-Coverage Challenge Corpus for
|
44 |
-
Sentence Understanding through Inference",
|
45 |
-
booktitle = "Proceedings of the 2018 Conference of
|
46 |
-
the North American Chapter of the
|
47 |
-
Association for Computational Linguistics:
|
48 |
-
Human Language Technologies, Volume 1 (Long
|
49 |
-
Papers)",
|
50 |
-
year = "2018",
|
51 |
-
publisher = "Association for Computational Linguistics",
|
52 |
-
pages = "1112--1122",
|
53 |
-
location = "New Orleans, Louisiana",
|
54 |
-
url = "http://aclweb.org/anthology/N18-1101"
|
55 |
-
}
|
56 |
-
@article{bowman2015large,
|
57 |
-
title={A large annotated corpus for learning natural language inference},
|
58 |
-
author={Bowman, Samuel R and Angeli, Gabor and Potts, Christopher and Manning, Christopher D},
|
59 |
-
journal={arXiv preprint arXiv:1508.05326},
|
60 |
-
year={2015}
|
61 |
-
}"""
|
62 |
-
),
|
63 |
-
url="http://www.nyu.edu/projects/bowman/multinli/",
|
64 |
-
)
|
65 |
-
|
66 |
-
ADVGLUE_DEV_URL = "https://adversarialglue.github.io/dataset/dev.zip"
|
67 |
-
|
68 |
-
|
69 |
-
class AdvGlueConfig(datasets.BuilderConfig):
|
70 |
-
"""BuilderConfig for Adversarial GLUE."""
|
71 |
-
|
72 |
-
def __init__(
|
73 |
-
self,
|
74 |
-
text_features,
|
75 |
-
label_column,
|
76 |
-
data_url,
|
77 |
-
data_dir,
|
78 |
-
citation,
|
79 |
-
url,
|
80 |
-
label_classes=None,
|
81 |
-
process_label=lambda x: x,
|
82 |
-
**kwargs,
|
83 |
-
):
|
84 |
-
"""BuilderConfig for Adversarial GLUE.
|
85 |
-
|
86 |
-
Args:
|
87 |
-
text_features: `dict[string, string]`, map from the name of the feature
|
88 |
-
dict for each text field to the name of the column in the tsv file
|
89 |
-
label_column: `string`, name of the column in the tsv file corresponding
|
90 |
-
to the label
|
91 |
-
data_url: `string`, url to download the zip file from
|
92 |
-
data_dir: `string`, the path to the folder containing the tsv files in the
|
93 |
-
downloaded zip
|
94 |
-
citation: `string`, citation for the data set
|
95 |
-
url: `string`, url for information about the data set
|
96 |
-
label_classes: `list[string]`, the list of classes if the label is
|
97 |
-
categorical. If not provided, then the label will be of type
|
98 |
-
`datasets.Value('float32')`.
|
99 |
-
process_label: `Function[string, any]`, function taking in the raw value
|
100 |
-
of the label and processing it to the form required by the label feature
|
101 |
-
**kwargs: keyword arguments forwarded to super.
|
102 |
-
"""
|
103 |
-
super(AdvGlueConfig, self).__init__(version=datasets.Version("1.0.0", ""), **kwargs)
|
104 |
-
self.text_features = text_features
|
105 |
-
self.label_column = label_column
|
106 |
-
self.label_classes = label_classes
|
107 |
-
self.data_url = data_url
|
108 |
-
self.data_dir = data_dir
|
109 |
-
self.citation = citation
|
110 |
-
self.url = url
|
111 |
-
self.process_label = process_label
|
112 |
-
|
113 |
-
|
114 |
-
ADVGLUE_BUILDER_CONFIGS = [
|
115 |
-
AdvGlueConfig(
|
116 |
-
name="adv_sst2",
|
117 |
-
description=textwrap.dedent(
|
118 |
-
"""Adversarial version of SST-2.
|
119 |
-
The Stanford Sentiment Treebank consists of sentences from movie reviews and
|
120 |
-
human annotations of their sentiment. The task is to predict the sentiment of a
|
121 |
-
given sentence. We use the two-way (positive/negative) class split, and use only
|
122 |
-
sentence-level labels."""
|
123 |
-
),
|
124 |
-
text_features={"sentence": "sentence"},
|
125 |
-
label_classes=["negative", "positive"],
|
126 |
-
label_column="label",
|
127 |
-
data_url="https://dl.fbaipublicfiles.com/glue/data/SST-2.zip",
|
128 |
-
data_dir="SST-2",
|
129 |
-
citation=textwrap.dedent(
|
130 |
-
"""\
|
131 |
-
@inproceedings{socher2013recursive,
|
132 |
-
title={Recursive deep models for semantic compositionality over a sentiment treebank},
|
133 |
-
author={Socher, Richard and Perelygin, Alex and Wu, Jean and Chuang, Jason and Manning, Christopher D and Ng, Andrew and Potts, Christopher},
|
134 |
-
booktitle={Proceedings of the 2013 conference on empirical methods in natural language processing},
|
135 |
-
pages={1631--1642},
|
136 |
-
year={2013}
|
137 |
-
}"""
|
138 |
-
),
|
139 |
-
url="https://datasets.stanford.edu/sentiment/index.html",
|
140 |
-
),
|
141 |
-
AdvGlueConfig(
|
142 |
-
name="adv_qqp",
|
143 |
-
description=textwrap.dedent(
|
144 |
-
"""Adversarial version of QQP.
|
145 |
-
The Quora Question Pairs2 dataset is a collection of question pairs from the
|
146 |
-
community question-answering website Quora. The task is to determine whether a
|
147 |
-
pair of questions are semantically equivalent."""
|
148 |
-
),
|
149 |
-
text_features={
|
150 |
-
"question1": "question1",
|
151 |
-
"question2": "question2",
|
152 |
-
},
|
153 |
-
label_classes=["not_duplicate", "duplicate"],
|
154 |
-
label_column="label",
|
155 |
-
data_url="https://dl.fbaipublicfiles.com/glue/data/QQP-clean.zip",
|
156 |
-
data_dir="QQP",
|
157 |
-
citation=textwrap.dedent(
|
158 |
-
"""\
|
159 |
-
@online{WinNT,
|
160 |
-
author = {Iyer, Shankar and Dandekar, Nikhil and Csernai, Kornel},
|
161 |
-
title = {First Quora Dataset Release: Question Pairs},
|
162 |
-
year = {2017},
|
163 |
-
url = {https://data.quora.com/First-Quora-Dataset-Release-Question-Pairs},
|
164 |
-
urldate = {2019-04-03}
|
165 |
-
}"""
|
166 |
-
),
|
167 |
-
url="https://data.quora.com/First-Quora-Dataset-Release-Question-Pairs",
|
168 |
-
),
|
169 |
-
AdvGlueConfig(
|
170 |
-
name="adv_mnli",
|
171 |
-
description=textwrap.dedent(
|
172 |
-
"""Adversarial version of MNLI.
|
173 |
-
The Multi-Genre Natural Language Inference Corpus is a crowdsourced
|
174 |
-
collection of sentence pairs with textual entailment annotations. Given a premise sentence
|
175 |
-
and a hypothesis sentence, the task is to predict whether the premise entails the hypothesis
|
176 |
-
(entailment), contradicts the hypothesis (contradiction), or neither (neutral). The premise sentences are
|
177 |
-
gathered from ten different sources, including transcribed speech, fiction, and government reports.
|
178 |
-
We use the standard test set, for which we obtained private labels from the authors, and evaluate
|
179 |
-
on both the matched (in-domain) and mismatched (cross-domain) section. We also use and recommend
|
180 |
-
the SNLI corpus as 550k examples of auxiliary training data."""
|
181 |
-
),
|
182 |
-
**_MNLI_BASE_KWARGS,
|
183 |
-
),
|
184 |
-
AdvGlueConfig(
|
185 |
-
name="adv_mnli_mismatched",
|
186 |
-
description=textwrap.dedent(
|
187 |
-
"""Adversarial version of MNLI-mismatched.
|
188 |
-
The mismatched validation and test splits from MNLI.
|
189 |
-
See the "mnli" BuilderConfig for additional information."""
|
190 |
-
),
|
191 |
-
**_MNLI_BASE_KWARGS,
|
192 |
-
),
|
193 |
-
AdvGlueConfig(
|
194 |
-
name="adv_qnli",
|
195 |
-
description=textwrap.dedent(
|
196 |
-
"""Adversarial version of QNLI.
|
197 |
-
The Stanford Question Answering Dataset is a question-answering
|
198 |
-
dataset consisting of question-paragraph pairs, where one of the sentences in the paragraph (drawn
|
199 |
-
from Wikipedia) contains the answer to the corresponding question (written by an annotator). We
|
200 |
-
convert the task into sentence pair classification by forming a pair between each question and each
|
201 |
-
sentence in the corresponding context, and filtering out pairs with low lexical overlap between the
|
202 |
-
question and the context sentence. The task is to determine whether the context sentence contains
|
203 |
-
the answer to the question. This modified version of the original task removes the requirement that
|
204 |
-
the model select the exact answer, but also removes the simplifying assumptions that the answer
|
205 |
-
is always present in the input and that lexical overlap is a reliable cue."""
|
206 |
-
), # pylint: disable=line-too-long
|
207 |
-
text_features={
|
208 |
-
"question": "question",
|
209 |
-
"sentence": "sentence",
|
210 |
-
},
|
211 |
-
label_classes=["entailment", "not_entailment"],
|
212 |
-
label_column="label",
|
213 |
-
data_url="https://dl.fbaipublicfiles.com/glue/data/QNLIv2.zip",
|
214 |
-
data_dir="QNLI",
|
215 |
-
citation=textwrap.dedent(
|
216 |
-
"""\
|
217 |
-
@article{rajpurkar2016squad,
|
218 |
-
title={Squad: 100,000+ questions for machine comprehension of text},
|
219 |
-
author={Rajpurkar, Pranav and Zhang, Jian and Lopyrev, Konstantin and Liang, Percy},
|
220 |
-
journal={arXiv preprint arXiv:1606.05250},
|
221 |
-
year={2016}
|
222 |
-
}"""
|
223 |
-
),
|
224 |
-
url="https://rajpurkar.github.io/SQuAD-explorer/",
|
225 |
-
),
|
226 |
-
AdvGlueConfig(
|
227 |
-
name="adv_rte",
|
228 |
-
description=textwrap.dedent(
|
229 |
-
"""Adversarial version of RTE.
|
230 |
-
The Recognizing Textual Entailment (RTE) datasets come from a series of annual textual
|
231 |
-
entailment challenges. We combine the data from RTE1 (Dagan et al., 2006), RTE2 (Bar Haim
|
232 |
-
et al., 2006), RTE3 (Giampiccolo et al., 2007), and RTE5 (Bentivogli et al., 2009).4 Examples are
|
233 |
-
constructed based on news and Wikipedia text. We convert all datasets to a two-class split, where
|
234 |
-
for three-class datasets we collapse neutral and contradiction into not entailment, for consistency."""
|
235 |
-
), # pylint: disable=line-too-long
|
236 |
-
text_features={
|
237 |
-
"sentence1": "sentence1",
|
238 |
-
"sentence2": "sentence2",
|
239 |
-
},
|
240 |
-
label_classes=["entailment", "not_entailment"],
|
241 |
-
label_column="label",
|
242 |
-
data_url="https://dl.fbaipublicfiles.com/glue/data/RTE.zip",
|
243 |
-
data_dir="RTE",
|
244 |
-
citation=textwrap.dedent(
|
245 |
-
"""\
|
246 |
-
@inproceedings{dagan2005pascal,
|
247 |
-
title={The PASCAL recognising textual entailment challenge},
|
248 |
-
author={Dagan, Ido and Glickman, Oren and Magnini, Bernardo},
|
249 |
-
booktitle={Machine Learning Challenges Workshop},
|
250 |
-
pages={177--190},
|
251 |
-
year={2005},
|
252 |
-
organization={Springer}
|
253 |
-
}
|
254 |
-
@inproceedings{bar2006second,
|
255 |
-
title={The second pascal recognising textual entailment challenge},
|
256 |
-
author={Bar-Haim, Roy and Dagan, Ido and Dolan, Bill and Ferro, Lisa and Giampiccolo, Danilo and Magnini, Bernardo and Szpektor, Idan},
|
257 |
-
booktitle={Proceedings of the second PASCAL challenges workshop on recognising textual entailment},
|
258 |
-
volume={6},
|
259 |
-
number={1},
|
260 |
-
pages={6--4},
|
261 |
-
year={2006},
|
262 |
-
organization={Venice}
|
263 |
-
}
|
264 |
-
@inproceedings{giampiccolo2007third,
|
265 |
-
title={The third pascal recognizing textual entailment challenge},
|
266 |
-
author={Giampiccolo, Danilo and Magnini, Bernardo and Dagan, Ido and Dolan, Bill},
|
267 |
-
booktitle={Proceedings of the ACL-PASCAL workshop on textual entailment and paraphrasing},
|
268 |
-
pages={1--9},
|
269 |
-
year={2007},
|
270 |
-
organization={Association for Computational Linguistics}
|
271 |
-
}
|
272 |
-
@inproceedings{bentivogli2009fifth,
|
273 |
-
title={The Fifth PASCAL Recognizing Textual Entailment Challenge.},
|
274 |
-
author={Bentivogli, Luisa and Clark, Peter and Dagan, Ido and Giampiccolo, Danilo},
|
275 |
-
booktitle={TAC},
|
276 |
-
year={2009}
|
277 |
-
}"""
|
278 |
-
),
|
279 |
-
url="https://aclweb.org/aclwiki/Recognizing_Textual_Entailment",
|
280 |
-
),
|
281 |
-
]
|
282 |
-
|
283 |
-
|
284 |
-
class AdvGlue(datasets.GeneratorBasedBuilder):
|
285 |
-
"""The General Language Understanding Evaluation (GLUE) benchmark."""
|
286 |
-
|
287 |
-
DATASETS = ["adv_sst2", "adv_qqp", "adv_mnli", "adv_mnli_mismatched", "adv_qnli", "adv_rte"]
|
288 |
-
BUILDER_CONFIGS = ADVGLUE_BUILDER_CONFIGS
|
289 |
-
|
290 |
-
def _info(self):
|
291 |
-
features = {text_feature: datasets.Value("string") for text_feature in self.config.text_features.keys()}
|
292 |
-
if self.config.label_classes:
|
293 |
-
features["label"] = datasets.features.ClassLabel(names=self.config.label_classes)
|
294 |
-
else:
|
295 |
-
features["label"] = datasets.Value("float32")
|
296 |
-
features["idx"] = datasets.Value("int32")
|
297 |
-
return datasets.DatasetInfo(
|
298 |
-
description=_ADV_GLUE_DESCRIPTION,
|
299 |
-
features=datasets.Features(features),
|
300 |
-
homepage="https://adversarialglue.github.io/",
|
301 |
-
citation=_ADV_GLUE_CITATION,
|
302 |
-
)
|
303 |
-
|
304 |
-
def _split_generators(self, dl_manager):
|
305 |
-
assert self.config.name in AdvGlue.DATASETS
|
306 |
-
data_dir = dl_manager.download_and_extract(ADVGLUE_DEV_URL)
|
307 |
-
data_file = os.path.join(data_dir, "dev", "dev.json")
|
308 |
-
return [
|
309 |
-
datasets.SplitGenerator(
|
310 |
-
name=datasets.Split.VALIDATION,
|
311 |
-
gen_kwargs={
|
312 |
-
"data_file": data_file,
|
313 |
-
},
|
314 |
-
)
|
315 |
-
]
|
316 |
-
|
317 |
-
def _generate_examples(self, data_file):
|
318 |
-
# We name splits 'adv_sst2' instead of 'sst2' so as not to be confused
|
319 |
-
# with the original SST-2. Here they're named like 'sst2' so we have to
|
320 |
-
# remove the 'adv_' prefix.
|
321 |
-
config_key = self.config.name.replace("adv_", "")
|
322 |
-
if config_key == "mnli_mismatched":
|
323 |
-
# and they name this split differently.
|
324 |
-
config_key = "mnli-mm"
|
325 |
-
data = json.loads(open(data_file).read())
|
326 |
-
for row in data[config_key]:
|
327 |
-
example = {feat: row[col] for feat, col in self.config.text_features.items()}
|
328 |
-
example["label"] = self.config.process_label(row[self.config.label_column])
|
329 |
-
example["idx"] = row["idx"]
|
330 |
-
yield example["idx"], example
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|