Hugging Face's logo

Datasets:
imppres

Tasks:
Text Classification
Sub-tasks:
natural-language-inference
Languages:
English
Multilinguality:
monolingual
Size Categories:
1K<n<10K
Language Creators:
machine-generated
Annotations Creators:
machine-generated
Source Datasets:
original
Tags:
License:
Dataset card Viewer Files Community
4
albertvillanova HF staff commited on
Commit
c99d162
•
1 Parent(s): 98a4184

Delete legacy dataset_infos.json

Browse files
Files changed (1) hide show
  1. dataset_infos.json +0 -1126
dataset_infos.json DELETED
@@ -1,1126 +0,0 @@
1
- {
2
- "presupposition_all_n_presupposition": {
3
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
4
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
5
- "homepage": "https://github.com/facebookresearch/Imppres",
6
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
7
- "features": {
8
- "premise": {
9
- "dtype": "string",
10
- "_type": "Value"
11
- },
12
- "hypothesis": {
13
- "dtype": "string",
14
- "_type": "Value"
15
- },
16
- "trigger": {
17
- "dtype": "string",
18
- "_type": "Value"
19
- },
20
- "trigger1": {
21
- "dtype": "string",
22
- "_type": "Value"
23
- },
24
- "trigger2": {
25
- "dtype": "string",
26
- "_type": "Value"
27
- },
28
- "presupposition": {
29
- "dtype": "string",
30
- "_type": "Value"
31
- },
32
- "gold_label": {
33
- "names": [
34
- "entailment",
35
- "neutral",
36
- "contradiction"
37
- ],
38
- "_type": "ClassLabel"
39
- },
40
- "UID": {
41
- "dtype": "string",
42
- "_type": "Value"
43
- },
44
- "pairID": {
45
- "dtype": "string",
46
- "_type": "Value"
47
- },
48
- "paradigmID": {
49
- "dtype": "int16",
50
- "_type": "Value"
51
- }
52
- },
53
- "builder_name": "parquet",
54
- "dataset_name": "imppres",
55
- "config_name": "presupposition_all_n_presupposition",
56
- "version": {
57
- "version_str": "1.1.0",
58
- "major": 1,
59
- "minor": 1,
60
- "patch": 0
61
- },
62
- "splits": {
63
- "all_n_presupposition": {
64
- "name": "all_n_presupposition",
65
- "num_bytes": 458460,
66
- "num_examples": 1900,
67
- "dataset_name": null
68
- }
69
- },
70
- "download_size": 43038,
71
- "dataset_size": 458460,
72
- "size_in_bytes": 501498
73
- },
74
- "presupposition_both_presupposition": {
75
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
76
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
77
- "homepage": "https://github.com/facebookresearch/Imppres",
78
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
79
- "features": {
80
- "premise": {
81
- "dtype": "string",
82
- "_type": "Value"
83
- },
84
- "hypothesis": {
85
- "dtype": "string",
86
- "_type": "Value"
87
- },
88
- "trigger": {
89
- "dtype": "string",
90
- "_type": "Value"
91
- },
92
- "trigger1": {
93
- "dtype": "string",
94
- "_type": "Value"
95
- },
96
- "trigger2": {
97
- "dtype": "string",
98
- "_type": "Value"
99
- },
100
- "presupposition": {
101
- "dtype": "string",
102
- "_type": "Value"
103
- },
104
- "gold_label": {
105
- "names": [
106
- "entailment",
107
- "neutral",
108
- "contradiction"
109
- ],
110
- "_type": "ClassLabel"
111
- },
112
- "UID": {
113
- "dtype": "string",
114
- "_type": "Value"
115
- },
116
- "pairID": {
117
- "dtype": "string",
118
- "_type": "Value"
119
- },
120
- "paradigmID": {
121
- "dtype": "int16",
122
- "_type": "Value"
123
- }
124
- },
125
- "builder_name": "parquet",
126
- "dataset_name": "imppres",
127
- "config_name": "presupposition_both_presupposition",
128
- "version": {
129
- "version_str": "1.1.0",
130
- "major": 1,
131
- "minor": 1,
132
- "patch": 0
133
- },
134
- "splits": {
135
- "both_presupposition": {
136
- "name": "both_presupposition",
137
- "num_bytes": 432760,
138
- "num_examples": 1900,
139
- "dataset_name": null
140
- }
141
- },
142
- "download_size": 41142,
143
- "dataset_size": 432760,
144
- "size_in_bytes": 473902
145
- },
146
- "presupposition_change_of_state": {
147
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
148
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
149
- "homepage": "https://github.com/facebookresearch/Imppres",
150
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
151
- "features": {
152
- "premise": {
153
- "dtype": "string",
154
- "_type": "Value"
155
- },
156
- "hypothesis": {
157
- "dtype": "string",
158
- "_type": "Value"
159
- },
160
- "trigger": {
161
- "dtype": "string",
162
- "_type": "Value"
163
- },
164
- "trigger1": {
165
- "dtype": "string",
166
- "_type": "Value"
167
- },
168
- "trigger2": {
169
- "dtype": "string",
170
- "_type": "Value"
171
- },
172
- "presupposition": {
173
- "dtype": "string",
174
- "_type": "Value"
175
- },
176
- "gold_label": {
177
- "names": [
178
- "entailment",
179
- "neutral",
180
- "contradiction"
181
- ],
182
- "_type": "ClassLabel"
183
- },
184
- "UID": {
185
- "dtype": "string",
186
- "_type": "Value"
187
- },
188
- "pairID": {
189
- "dtype": "string",
190
- "_type": "Value"
191
- },
192
- "paradigmID": {
193
- "dtype": "int16",
194
- "_type": "Value"
195
- }
196
- },
197
- "builder_name": "parquet",
198
- "dataset_name": "imppres",
199
- "config_name": "presupposition_change_of_state",
200
- "version": {
201
- "version_str": "1.1.0",
202
- "major": 1,
203
- "minor": 1,
204
- "patch": 0
205
- },
206
- "splits": {
207
- "change_of_state": {
208
- "name": "change_of_state",
209
- "num_bytes": 308595,
210
- "num_examples": 1900,
211
- "dataset_name": null
212
- }
213
- },
214
- "download_size": 35814,
215
- "dataset_size": 308595,
216
- "size_in_bytes": 344409
217
- },
218
- "presupposition_cleft_existence": {
219
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
220
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
221
- "homepage": "https://github.com/facebookresearch/Imppres",
222
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
223
- "features": {
224
- "premise": {
225
- "dtype": "string",
226
- "_type": "Value"
227
- },
228
- "hypothesis": {
229
- "dtype": "string",
230
- "_type": "Value"
231
- },
232
- "trigger": {
233
- "dtype": "string",
234
- "_type": "Value"
235
- },
236
- "trigger1": {
237
- "dtype": "string",
238
- "_type": "Value"
239
- },
240
- "trigger2": {
241
- "dtype": "string",
242
- "_type": "Value"
243
- },
244
- "presupposition": {
245
- "dtype": "string",
246
- "_type": "Value"
247
- },
248
- "gold_label": {
249
- "names": [
250
- "entailment",
251
- "neutral",
252
- "contradiction"
253
- ],
254
- "_type": "ClassLabel"
255
- },
256
- "UID": {
257
- "dtype": "string",
258
- "_type": "Value"
259
- },
260
- "pairID": {
261
- "dtype": "string",
262
- "_type": "Value"
263
- },
264
- "paradigmID": {
265
- "dtype": "int16",
266
- "_type": "Value"
267
- }
268
- },
269
- "builder_name": "parquet",
270
- "dataset_name": "imppres",
271
- "config_name": "presupposition_cleft_existence",
272
- "version": {
273
- "version_str": "1.1.0",
274
- "major": 1,
275
- "minor": 1,
276
- "patch": 0
277
- },
278
- "splits": {
279
- "cleft_existence": {
280
- "name": "cleft_existence",
281
- "num_bytes": 363206,
282
- "num_examples": 1900,
283
- "dataset_name": null
284
- }
285
- },
286
- "download_size": 37597,
287
- "dataset_size": 363206,
288
- "size_in_bytes": 400803
289
- },
290
- "presupposition_cleft_uniqueness": {
291
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
292
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
293
- "homepage": "https://github.com/facebookresearch/Imppres",
294
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
295
- "features": {
296
- "premise": {
297
- "dtype": "string",
298
- "_type": "Value"
299
- },
300
- "hypothesis": {
301
- "dtype": "string",
302
- "_type": "Value"
303
- },
304
- "trigger": {
305
- "dtype": "string",
306
- "_type": "Value"
307
- },
308
- "trigger1": {
309
- "dtype": "string",
310
- "_type": "Value"
311
- },
312
- "trigger2": {
313
- "dtype": "string",
314
- "_type": "Value"
315
- },
316
- "presupposition": {
317
- "dtype": "string",
318
- "_type": "Value"
319
- },
320
- "gold_label": {
321
- "names": [
322
- "entailment",
323
- "neutral",
324
- "contradiction"
325
- ],
326
- "_type": "ClassLabel"
327
- },
328
- "UID": {
329
- "dtype": "string",
330
- "_type": "Value"
331
- },
332
- "pairID": {
333
- "dtype": "string",
334
- "_type": "Value"
335
- },
336
- "paradigmID": {
337
- "dtype": "int16",
338
- "_type": "Value"
339
- }
340
- },
341
- "builder_name": "parquet",
342
- "dataset_name": "imppres",
343
- "config_name": "presupposition_cleft_uniqueness",
344
- "version": {
345
- "version_str": "1.1.0",
346
- "major": 1,
347
- "minor": 1,
348
- "patch": 0
349
- },
350
- "splits": {
351
- "cleft_uniqueness": {
352
- "name": "cleft_uniqueness",
353
- "num_bytes": 388747,
354
- "num_examples": 1900,
355
- "dataset_name": null
356
- }
357
- },
358
- "download_size": 38279,
359
- "dataset_size": 388747,
360
- "size_in_bytes": 427026
361
- },
362
- "presupposition_only_presupposition": {
363
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
364
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
365
- "homepage": "https://github.com/facebookresearch/Imppres",
366
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
367
- "features": {
368
- "premise": {
369
- "dtype": "string",
370
- "_type": "Value"
371
- },
372
- "hypothesis": {
373
- "dtype": "string",
374
- "_type": "Value"
375
- },
376
- "trigger": {
377
- "dtype": "string",
378
- "_type": "Value"
379
- },
380
- "trigger1": {
381
- "dtype": "string",
382
- "_type": "Value"
383
- },
384
- "trigger2": {
385
- "dtype": "string",
386
- "_type": "Value"
387
- },
388
- "presupposition": {
389
- "dtype": "string",
390
- "_type": "Value"
391
- },
392
- "gold_label": {
393
- "names": [
394
- "entailment",
395
- "neutral",
396
- "contradiction"
397
- ],
398
- "_type": "ClassLabel"
399
- },
400
- "UID": {
401
- "dtype": "string",
402
- "_type": "Value"
403
- },
404
- "pairID": {
405
- "dtype": "string",
406
- "_type": "Value"
407
- },
408
- "paradigmID": {
409
- "dtype": "int16",
410
- "_type": "Value"
411
- }
412
- },
413
- "builder_name": "parquet",
414
- "dataset_name": "imppres",
415
- "config_name": "presupposition_only_presupposition",
416
- "version": {
417
- "version_str": "1.1.0",
418
- "major": 1,
419
- "minor": 1,
420
- "patch": 0
421
- },
422
- "splits": {
423
- "only_presupposition": {
424
- "name": "only_presupposition",
425
- "num_bytes": 348986,
426
- "num_examples": 1900,
427
- "dataset_name": null
428
- }
429
- },
430
- "download_size": 38126,
431
- "dataset_size": 348986,
432
- "size_in_bytes": 387112
433
- },
434
- "presupposition_possessed_definites_existence": {
435
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
436
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
437
- "homepage": "https://github.com/facebookresearch/Imppres",
438
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
439
- "features": {
440
- "premise": {
441
- "dtype": "string",
442
- "_type": "Value"
443
- },
444
- "hypothesis": {
445
- "dtype": "string",
446
- "_type": "Value"
447
- },
448
- "trigger": {
449
- "dtype": "string",
450
- "_type": "Value"
451
- },
452
- "trigger1": {
453
- "dtype": "string",
454
- "_type": "Value"
455
- },
456
- "trigger2": {
457
- "dtype": "string",
458
- "_type": "Value"
459
- },
460
- "presupposition": {
461
- "dtype": "string",
462
- "_type": "Value"
463
- },
464
- "gold_label": {
465
- "names": [
466
- "entailment",
467
- "neutral",
468
- "contradiction"
469
- ],
470
- "_type": "ClassLabel"
471
- },
472
- "UID": {
473
- "dtype": "string",
474
- "_type": "Value"
475
- },
476
- "pairID": {
477
- "dtype": "string",
478
- "_type": "Value"
479
- },
480
- "paradigmID": {
481
- "dtype": "int16",
482
- "_type": "Value"
483
- }
484
- },
485
- "builder_name": "parquet",
486
- "dataset_name": "imppres",
487
- "config_name": "presupposition_possessed_definites_existence",
488
- "version": {
489
- "version_str": "1.1.0",
490
- "major": 1,
491
- "minor": 1,
492
- "patch": 0
493
- },
494
- "splits": {
495
- "possessed_definites_existence": {
496
- "name": "possessed_definites_existence",
497
- "num_bytes": 362302,
498
- "num_examples": 1900,
499
- "dataset_name": null
500
- }
501
- },
502
- "download_size": 38712,
503
- "dataset_size": 362302,
504
- "size_in_bytes": 401014
505
- },
506
- "presupposition_possessed_definites_uniqueness": {
507
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
508
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
509
- "homepage": "https://github.com/facebookresearch/Imppres",
510
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
511
- "features": {
512
- "premise": {
513
- "dtype": "string",
514
- "_type": "Value"
515
- },
516
- "hypothesis": {
517
- "dtype": "string",
518
- "_type": "Value"
519
- },
520
- "trigger": {
521
- "dtype": "string",
522
- "_type": "Value"
523
- },
524
- "trigger1": {
525
- "dtype": "string",
526
- "_type": "Value"
527
- },
528
- "trigger2": {
529
- "dtype": "string",
530
- "_type": "Value"
531
- },
532
- "presupposition": {
533
- "dtype": "string",
534
- "_type": "Value"
535
- },
536
- "gold_label": {
537
- "names": [
538
- "entailment",
539
- "neutral",
540
- "contradiction"
541
- ],
542
- "_type": "ClassLabel"
543
- },
544
- "UID": {
545
- "dtype": "string",
546
- "_type": "Value"
547
- },
548
- "pairID": {
549
- "dtype": "string",
550
- "_type": "Value"
551
- },
552
- "paradigmID": {
553
- "dtype": "int16",
554
- "_type": "Value"
555
- }
556
- },
557
- "builder_name": "parquet",
558
- "dataset_name": "imppres",
559
- "config_name": "presupposition_possessed_definites_uniqueness",
560
- "version": {
561
- "version_str": "1.1.0",
562
- "major": 1,
563
- "minor": 1,
564
- "patch": 0
565
- },
566
- "splits": {
567
- "possessed_definites_uniqueness": {
568
- "name": "possessed_definites_uniqueness",
569
- "num_bytes": 459371,
570
- "num_examples": 1900,
571
- "dataset_name": null
572
- }
573
- },
574
- "download_size": 42068,
575
- "dataset_size": 459371,
576
- "size_in_bytes": 501439
577
- },
578
- "presupposition_question_presupposition": {
579
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
580
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
581
- "homepage": "https://github.com/facebookresearch/Imppres",
582
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
583
- "features": {
584
- "premise": {
585
- "dtype": "string",
586
- "_type": "Value"
587
- },
588
- "hypothesis": {
589
- "dtype": "string",
590
- "_type": "Value"
591
- },
592
- "trigger": {
593
- "dtype": "string",
594
- "_type": "Value"
595
- },
596
- "trigger1": {
597
- "dtype": "string",
598
- "_type": "Value"
599
- },
600
- "trigger2": {
601
- "dtype": "string",
602
- "_type": "Value"
603
- },
604
- "presupposition": {
605
- "dtype": "string",
606
- "_type": "Value"
607
- },
608
- "gold_label": {
609
- "names": [
610
- "entailment",
611
- "neutral",
612
- "contradiction"
613
- ],
614
- "_type": "ClassLabel"
615
- },
616
- "UID": {
617
- "dtype": "string",
618
- "_type": "Value"
619
- },
620
- "pairID": {
621
- "dtype": "string",
622
- "_type": "Value"
623
- },
624
- "paradigmID": {
625
- "dtype": "int16",
626
- "_type": "Value"
627
- }
628
- },
629
- "builder_name": "parquet",
630
- "dataset_name": "imppres",
631
- "config_name": "presupposition_question_presupposition",
632
- "version": {
633
- "version_str": "1.1.0",
634
- "major": 1,
635
- "minor": 1,
636
- "patch": 0
637
- },
638
- "splits": {
639
- "question_presupposition": {
640
- "name": "question_presupposition",
641
- "num_bytes": 397195,
642
- "num_examples": 1900,
643
- "dataset_name": null
644
- }
645
- },
646
- "download_size": 41247,
647
- "dataset_size": 397195,
648
- "size_in_bytes": 438442
649
- },
650
- "implicature_connectives": {
651
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
652
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
653
- "homepage": "https://github.com/facebookresearch/Imppres",
654
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
655
- "features": {
656
- "premise": {
657
- "dtype": "string",
658
- "_type": "Value"
659
- },
660
- "hypothesis": {
661
- "dtype": "string",
662
- "_type": "Value"
663
- },
664
- "gold_label_log": {
665
- "names": [
666
- "entailment",
667
- "neutral",
668
- "contradiction"
669
- ],
670
- "_type": "ClassLabel"
671
- },
672
- "gold_label_prag": {
673
- "names": [
674
- "entailment",
675
- "neutral",
676
- "contradiction"
677
- ],
678
- "_type": "ClassLabel"
679
- },
680
- "spec_relation": {
681
- "dtype": "string",
682
- "_type": "Value"
683
- },
684
- "item_type": {
685
- "dtype": "string",
686
- "_type": "Value"
687
- },
688
- "trigger": {
689
- "dtype": "string",
690
- "_type": "Value"
691
- },
692
- "lexemes": {
693
- "dtype": "string",
694
- "_type": "Value"
695
- }
696
- },
697
- "builder_name": "parquet",
698
- "dataset_name": "imppres",
699
- "config_name": "implicature_connectives",
700
- "version": {
701
- "version_str": "1.1.0",
702
- "major": 1,
703
- "minor": 1,
704
- "patch": 0
705
- },
706
- "splits": {
707
- "connectives": {
708
- "name": "connectives",
709
- "num_bytes": 221844,
710
- "num_examples": 1200,
711
- "dataset_name": null
712
- }
713
- },
714
- "download_size": 25478,
715
- "dataset_size": 221844,
716
- "size_in_bytes": 247322
717
- },
718
- "implicature_gradable_adjective": {
719
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
720
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
721
- "homepage": "https://github.com/facebookresearch/Imppres",
722
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
723
- "features": {
724
- "premise": {
725
- "dtype": "string",
726
- "_type": "Value"
727
- },
728
- "hypothesis": {
729
- "dtype": "string",
730
- "_type": "Value"
731
- },
732
- "gold_label_log": {
733
- "names": [
734
- "entailment",
735
- "neutral",
736
- "contradiction"
737
- ],
738
- "_type": "ClassLabel"
739
- },
740
- "gold_label_prag": {
741
- "names": [
742
- "entailment",
743
- "neutral",
744
- "contradiction"
745
- ],
746
- "_type": "ClassLabel"
747
- },
748
- "spec_relation": {
749
- "dtype": "string",
750
- "_type": "Value"
751
- },
752
- "item_type": {
753
- "dtype": "string",
754
- "_type": "Value"
755
- },
756
- "trigger": {
757
- "dtype": "string",
758
- "_type": "Value"
759
- },
760
- "lexemes": {
761
- "dtype": "string",
762
- "_type": "Value"
763
- }
764
- },
765
- "builder_name": "parquet",
766
- "dataset_name": "imppres",
767
- "config_name": "implicature_gradable_adjective",
768
- "version": {
769
- "version_str": "1.1.0",
770
- "major": 1,
771
- "minor": 1,
772
- "patch": 0
773
- },
774
- "splits": {
775
- "gradable_adjective": {
776
- "name": "gradable_adjective",
777
- "num_bytes": 153648,
778
- "num_examples": 1200,
779
- "dataset_name": null
780
- }
781
- },
782
- "download_size": 17337,
783
- "dataset_size": 153648,
784
- "size_in_bytes": 170985
785
- },
786
- "implicature_gradable_verb": {
787
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
788
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
789
- "homepage": "https://github.com/facebookresearch/Imppres",
790
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
791
- "features": {
792
- "premise": {
793
- "dtype": "string",
794
- "_type": "Value"
795
- },
796
- "hypothesis": {
797
- "dtype": "string",
798
- "_type": "Value"
799
- },
800
- "gold_label_log": {
801
- "names": [
802
- "entailment",
803
- "neutral",
804
- "contradiction"
805
- ],
806
- "_type": "ClassLabel"
807
- },
808
- "gold_label_prag": {
809
- "names": [
810
- "entailment",
811
- "neutral",
812
- "contradiction"
813
- ],
814
- "_type": "ClassLabel"
815
- },
816
- "spec_relation": {
817
- "dtype": "string",
818
- "_type": "Value"
819
- },
820
- "item_type": {
821
- "dtype": "string",
822
- "_type": "Value"
823
- },
824
- "trigger": {
825
- "dtype": "string",
826
- "_type": "Value"
827
- },
828
- "lexemes": {
829
- "dtype": "string",
830
- "_type": "Value"
831
- }
832
- },
833
- "builder_name": "parquet",
834
- "dataset_name": "imppres",
835
- "config_name": "implicature_gradable_verb",
836
- "version": {
837
- "version_str": "1.1.0",
838
- "major": 1,
839
- "minor": 1,
840
- "patch": 0
841
- },
842
- "splits": {
843
- "gradable_verb": {
844
- "name": "gradable_verb",
845
- "num_bytes": 180678,
846
- "num_examples": 1200,
847
- "dataset_name": null
848
- }
849
- },
850
- "download_size": 21504,
851
- "dataset_size": 180678,
852
- "size_in_bytes": 202182
853
- },
854
- "implicature_modals": {
855
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
856
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
857
- "homepage": "https://github.com/facebookresearch/Imppres",
858
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
859
- "features": {
860
- "premise": {
861
- "dtype": "string",
862
- "_type": "Value"
863
- },
864
- "hypothesis": {
865
- "dtype": "string",
866
- "_type": "Value"
867
- },
868
- "gold_label_log": {
869
- "names": [
870
- "entailment",
871
- "neutral",
872
- "contradiction"
873
- ],
874
- "_type": "ClassLabel"
875
- },
876
- "gold_label_prag": {
877
- "names": [
878
- "entailment",
879
- "neutral",
880
- "contradiction"
881
- ],
882
- "_type": "ClassLabel"
883
- },
884
- "spec_relation": {
885
- "dtype": "string",
886
- "_type": "Value"
887
- },
888
- "item_type": {
889
- "dtype": "string",
890
- "_type": "Value"
891
- },
892
- "trigger": {
893
- "dtype": "string",
894
- "_type": "Value"
895
- },
896
- "lexemes": {
897
- "dtype": "string",
898
- "_type": "Value"
899
- }
900
- },
901
- "builder_name": "parquet",
902
- "dataset_name": "imppres",
903
- "config_name": "implicature_modals",
904
- "version": {
905
- "version_str": "1.1.0",
906
- "major": 1,
907
- "minor": 1,
908
- "patch": 0
909
- },
910
- "splits": {
911
- "modals": {
912
- "name": "modals",
913
- "num_bytes": 178536,
914
- "num_examples": 1200,
915
- "dataset_name": null
916
- }
917
- },
918
- "download_size": 21179,
919
- "dataset_size": 178536,
920
- "size_in_bytes": 199715
921
- },
922
- "implicature_numerals_10_100": {
923
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
924
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
925
- "homepage": "https://github.com/facebookresearch/Imppres",
926
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
927
- "features": {
928
- "premise": {
929
- "dtype": "string",
930
- "_type": "Value"
931
- },
932
- "hypothesis": {
933
- "dtype": "string",
934
- "_type": "Value"
935
- },
936
- "gold_label_log": {
937
- "names": [
938
- "entailment",
939
- "neutral",
940
- "contradiction"
941
- ],
942
- "_type": "ClassLabel"
943
- },
944
- "gold_label_prag": {
945
- "names": [
946
- "entailment",
947
- "neutral",
948
- "contradiction"
949
- ],
950
- "_type": "ClassLabel"
951
- },
952
- "spec_relation": {
953
- "dtype": "string",
954
- "_type": "Value"
955
- },
956
- "item_type": {
957
- "dtype": "string",
958
- "_type": "Value"
959
- },
960
- "trigger": {
961
- "dtype": "string",
962
- "_type": "Value"
963
- },
964
- "lexemes": {
965
- "dtype": "string",
966
- "_type": "Value"
967
- }
968
- },
969
- "builder_name": "parquet",
970
- "dataset_name": "imppres",
971
- "config_name": "implicature_numerals_10_100",
972
- "version": {
973
- "version_str": "1.1.0",
974
- "major": 1,
975
- "minor": 1,
976
- "patch": 0
977
- },
978
- "splits": {
979
- "numerals_10_100": {
980
- "name": "numerals_10_100",
981
- "num_bytes": 208596,
982
- "num_examples": 1200,
983
- "dataset_name": null
984
- }
985
- },
986
- "download_size": 22640,
987
- "dataset_size": 208596,
988
- "size_in_bytes": 231236
989
- },
990
- "implicature_numerals_2_3": {
991
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
992
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
993
- "homepage": "https://github.com/facebookresearch/Imppres",
994
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
995
- "features": {
996
- "premise": {
997
- "dtype": "string",
998
- "_type": "Value"
999
- },
1000
- "hypothesis": {
1001
- "dtype": "string",
1002
- "_type": "Value"
1003
- },
1004
- "gold_label_log": {
1005
- "names": [
1006
- "entailment",
1007
- "neutral",
1008
- "contradiction"
1009
- ],
1010
- "_type": "ClassLabel"
1011
- },
1012
- "gold_label_prag": {
1013
- "names": [
1014
- "entailment",
1015
- "neutral",
1016
- "contradiction"
1017
- ],
1018
- "_type": "ClassLabel"
1019
- },
1020
- "spec_relation": {
1021
- "dtype": "string",
1022
- "_type": "Value"
1023
- },
1024
- "item_type": {
1025
- "dtype": "string",
1026
- "_type": "Value"
1027
- },
1028
- "trigger": {
1029
- "dtype": "string",
1030
- "_type": "Value"
1031
- },
1032
- "lexemes": {
1033
- "dtype": "string",
1034
- "_type": "Value"
1035
- }
1036
- },
1037
- "builder_name": "parquet",
1038
- "dataset_name": "imppres",
1039
- "config_name": "implicature_numerals_2_3",
1040
- "version": {
1041
- "version_str": "1.1.0",
1042
- "major": 1,
1043
- "minor": 1,
1044
- "patch": 0
1045
- },
1046
- "splits": {
1047
- "numerals_2_3": {
1048
- "name": "numerals_2_3",
1049
- "num_bytes": 188760,
1050
- "num_examples": 1200,
1051
- "dataset_name": null
1052
- }
1053
- },
1054
- "download_size": 22218,
1055
- "dataset_size": 188760,
1056
- "size_in_bytes": 210978
1057
- },
1058
- "implicature_quantifiers": {
1059
- "description": "Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.",
1060
- "citation": "@inproceedings{jeretic-etal-2020-natural,\n title = \"Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}\",\n author = \"Jereti\u000b{c}, Paloma and\n Warstadt, Alex and\n Bhooshan, Suvrat and\n Williams, Adina\",\n booktitle = \"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\",\n month = jul,\n year = \"2020\",\n address = \"Online\",\n publisher = \"Association for Computational Linguistics\",\n url = \"https://www.aclweb.org/anthology/2020.acl-main.768\",\n doi = \"10.18653/v1/2020.acl-main.768\",\n pages = \"8690--8705\",\n abstract = \"Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by {``}some{''} as entailments. For some presupposition triggers like {``}only{''}, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences.\",\n}\n",
1061
- "homepage": "https://github.com/facebookresearch/Imppres",
1062
- "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License",
1063
- "features": {
1064
- "premise": {
1065
- "dtype": "string",
1066
- "_type": "Value"
1067
- },
1068
- "hypothesis": {
1069
- "dtype": "string",
1070
- "_type": "Value"
1071
- },
1072
- "gold_label_log": {
1073
- "names": [
1074
- "entailment",
1075
- "neutral",
1076
- "contradiction"
1077
- ],
1078
- "_type": "ClassLabel"
1079
- },
1080
- "gold_label_prag": {
1081
- "names": [
1082
- "entailment",
1083
- "neutral",
1084
- "contradiction"
1085
- ],
1086
- "_type": "ClassLabel"
1087
- },
1088
- "spec_relation": {
1089
- "dtype": "string",
1090
- "_type": "Value"
1091
- },
1092
- "item_type": {
1093
- "dtype": "string",
1094
- "_type": "Value"
1095
- },
1096
- "trigger": {
1097
- "dtype": "string",
1098
- "_type": "Value"
1099
- },
1100
- "lexemes": {
1101
- "dtype": "string",
1102
- "_type": "Value"
1103
- }
1104
- },
1105
- "builder_name": "parquet",
1106
- "dataset_name": "imppres",
1107
- "config_name": "implicature_quantifiers",
1108
- "version": {
1109
- "version_str": "1.1.0",
1110
- "major": 1,
1111
- "minor": 1,
1112
- "patch": 0
1113
- },
1114
- "splits": {
1115
- "quantifiers": {
1116
- "name": "quantifiers",
1117
- "num_bytes": 176790,
1118
- "num_examples": 1200,
1119
- "dataset_name": null
1120
- }
1121
- },
1122
- "download_size": 21017,
1123
- "dataset_size": 176790,
1124
- "size_in_bytes": 197807
1125
- }
1126
- }