Datasets:
Tasks:
Text Classification
Modalities:
Text
Formats:
parquet
Sub-tasks:
natural-language-inference
Languages:
English
Size:
10K<n<100K
Tags:
License:
Commit
•
4490772
0
Parent(s):
Update files from the datasets library (from 1.2.0)
Browse filesRelease notes: https://github.com/huggingface/datasets/releases/tag/1.2.0
- .gitattributes +27 -0
- README.md +286 -0
- dataset_infos.json +1 -0
- dummy/implicature_connectives/1.1.0/dummy_data.zip +3 -0
- dummy/implicature_gradable_adjective/1.1.0/dummy_data.zip +3 -0
- dummy/implicature_gradable_verb/1.1.0/dummy_data.zip +3 -0
- dummy/implicature_modals/1.1.0/dummy_data.zip +3 -0
- dummy/implicature_numerals_10_100/1.1.0/dummy_data.zip +3 -0
- dummy/implicature_numerals_2_3/1.1.0/dummy_data.zip +3 -0
- dummy/implicature_quantifiers/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_all_n_presupposition/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_both_presupposition/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_change_of_state/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_cleft_existence/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_cleft_uniqueness/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_only_presupposition/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_possessed_definites_existence/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_possessed_definites_uniqueness/1.1.0/dummy_data.zip +3 -0
- dummy/presupposition_question_presupposition/1.1.0/dummy_data.zip +3 -0
- imppres.py +278 -0
.gitattributes
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
| 2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
| 3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
| 4 |
+
*.bin.* filter=lfs diff=lfs merge=lfs -text
|
| 5 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
| 6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
| 7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
| 8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
| 9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
| 10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
| 11 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
| 12 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
| 13 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
| 14 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
| 15 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
| 16 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
| 17 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
| 18 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
| 19 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
| 20 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
| 21 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
| 22 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
| 23 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
| 24 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
| 25 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 26 |
+
*.zstandard filter=lfs diff=lfs merge=lfs -text
|
| 27 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,286 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
annotations_creators:
|
| 3 |
+
- machine-generated
|
| 4 |
+
language_creators:
|
| 5 |
+
- machine-generated
|
| 6 |
+
languages:
|
| 7 |
+
- en
|
| 8 |
+
licenses:
|
| 9 |
+
- cc-by-nc-4-0
|
| 10 |
+
multilinguality:
|
| 11 |
+
- monolingual
|
| 12 |
+
size_categories:
|
| 13 |
+
- 10K<n<100K
|
| 14 |
+
source_datasets:
|
| 15 |
+
- original
|
| 16 |
+
task_categories:
|
| 17 |
+
- text-classification
|
| 18 |
+
task_ids:
|
| 19 |
+
- natural-language-inference
|
| 20 |
+
---
|
| 21 |
+
|
| 22 |
+
# Dataset Card for IMPPRES
|
| 23 |
+
|
| 24 |
+
## Table of Contents
|
| 25 |
+
- [Dataset Description](#dataset-description)
|
| 26 |
+
- [Dataset Summary](#dataset-summary)
|
| 27 |
+
- [Supported Tasks](#supported-tasks-and-leaderboards)
|
| 28 |
+
- [Languages](#languages)
|
| 29 |
+
- [Dataset Structure](#dataset-structure)
|
| 30 |
+
- [Data Instances](#data-instances)
|
| 31 |
+
- [Data Fields](#data-instances)
|
| 32 |
+
- [Data Splits](#data-instances)
|
| 33 |
+
- [Dataset Creation](#dataset-creation)
|
| 34 |
+
- [Curation Rationale](#curation-rationale)
|
| 35 |
+
- [Source Data](#source-data)
|
| 36 |
+
- [Annotations](#annotations)
|
| 37 |
+
- [Personal and Sensitive Information](#personal-and-sensitive-information)
|
| 38 |
+
- [Considerations for Using the Data](#considerations-for-using-the-data)
|
| 39 |
+
- [Social Impact of Dataset](#social-impact-of-dataset)
|
| 40 |
+
- [Discussion of Biases](#discussion-of-biases)
|
| 41 |
+
- [Other Known Limitations](#other-known-limitations)
|
| 42 |
+
- [Additional Information](#additional-information)
|
| 43 |
+
- [Dataset Curators](#dataset-curators)
|
| 44 |
+
- [Licensing Information](#licensing-information)
|
| 45 |
+
- [Citation Information](#citation-information)
|
| 46 |
+
|
| 47 |
+
## Dataset Description
|
| 48 |
+
|
| 49 |
+
- **Homepage:** [Github](https://github.com/facebookresearch/Imppres)
|
| 50 |
+
- **Repository:** [Github](https://github.com/facebookresearch/Imppres)
|
| 51 |
+
- **Paper:** [Aclweb](https://www.aclweb.org/anthology/2020.acl-main.768)
|
| 52 |
+
- **Leaderboard:**
|
| 53 |
+
- **Point of Contact:**
|
| 54 |
+
|
| 55 |
+
### Dataset Summary
|
| 56 |
+
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.
|
| 57 |
+
|
| 58 |
+
### Supported Tasks and Leaderboards
|
| 59 |
+
|
| 60 |
+
Natural Language Inference.
|
| 61 |
+
|
| 62 |
+
### Languages
|
| 63 |
+
|
| 64 |
+
English.
|
| 65 |
+
|
| 66 |
+
## Dataset Structure
|
| 67 |
+
|
| 68 |
+
### Data Instances
|
| 69 |
+
|
| 70 |
+
The data consists of 2 configurations: implicature and presupposition.
|
| 71 |
+
Each configuration consists of several different sub-datasets:
|
| 72 |
+
|
| 73 |
+
**Pressupposition**
|
| 74 |
+
- all_n_presupposition
|
| 75 |
+
- change_of_state
|
| 76 |
+
- cleft_uniqueness
|
| 77 |
+
- possessed_definites_existence
|
| 78 |
+
- question_presupposition
|
| 79 |
+
- both_presupposition
|
| 80 |
+
- cleft_existence
|
| 81 |
+
- only_presupposition
|
| 82 |
+
- possessed_definites_uniqueness
|
| 83 |
+
|
| 84 |
+
**Implicature**
|
| 85 |
+
- connectives
|
| 86 |
+
- gradable_adjective
|
| 87 |
+
- gradable_verb
|
| 88 |
+
- modals
|
| 89 |
+
- numerals_10_100
|
| 90 |
+
- numerals_2_3
|
| 91 |
+
- quantifiers
|
| 92 |
+
|
| 93 |
+
Each sentence type in IMPPRES is generated according to a template that specifies the linear order of the constituents in the sentence. The constituents are sampled from a vocabulary of over 3000 lexical items annotated with grammatical features needed to ensure wellformedness. We semiautomatically generate IMPPRES using a codebase developed by Warstadt et al. (2019a) and significantly expanded for the BLiMP dataset (Warstadt et al., 2019b).
|
| 94 |
+
|
| 95 |
+
Here is an instance of the raw presupposition data from any sub-dataset:
|
| 96 |
+
```buildoutcfg
|
| 97 |
+
{
|
| 98 |
+
"sentence1": "All ten guys that proved to boast might have been divorcing.",
|
| 99 |
+
"sentence2": "There are exactly ten guys that proved to boast.",
|
| 100 |
+
"trigger": "modal",
|
| 101 |
+
"presupposition": "positive",
|
| 102 |
+
"gold_label": "entailment",
|
| 103 |
+
"UID": "all_n_presupposition",
|
| 104 |
+
"pairID": "9e",
|
| 105 |
+
"paradigmID": 0
|
| 106 |
+
}
|
| 107 |
+
```
|
| 108 |
+
and the raw implicature data from any sub-dataset:
|
| 109 |
+
```buildoutcfg
|
| 110 |
+
{
|
| 111 |
+
"sentence1": "That teenager couldn't yell.",
|
| 112 |
+
"sentence2": "That teenager could yell.",
|
| 113 |
+
"gold_label_log": "contradiction",
|
| 114 |
+
"gold_label_prag": "contradiction",
|
| 115 |
+
"spec_relation": "negation",
|
| 116 |
+
"item_type": "control",
|
| 117 |
+
"trigger": "modal",
|
| 118 |
+
"lexemes": "can - have to"
|
| 119 |
+
}
|
| 120 |
+
```
|
| 121 |
+
### Data Fields
|
| 122 |
+
**Presupposition**
|
| 123 |
+
|
| 124 |
+
There is a slight mapping from the raw data fields in the presupposition sub-datasets and the fields appearing in the HuggingFace Datasets.
|
| 125 |
+
When dealing with the HF Dataset, the following mapping of fields happens:
|
| 126 |
+
```buildoutcfg
|
| 127 |
+
"premise" -> "sentence1"
|
| 128 |
+
"hypothesis"-> "sentence2"
|
| 129 |
+
"trigger" -> "trigger" or "Not_In_Example"
|
| 130 |
+
"trigger1" -> "trigger1" or "Not_In_Example"
|
| 131 |
+
"trigger2" -> "trigger2" or "Not_In_Example"
|
| 132 |
+
"presupposition" -> "presupposition" or "Not_In_Example"
|
| 133 |
+
"gold_label" -> "gold_label"
|
| 134 |
+
"UID" -> "UID"
|
| 135 |
+
"pairID" -> "pairID"
|
| 136 |
+
"paradigmID" -> "paradigmID"
|
| 137 |
+
```
|
| 138 |
+
For the most part, the majority of the raw fields remain unchanged. However, when it comes to the various `trigger` fields, a new mapping was introduced.
|
| 139 |
+
There are some examples in the dataset that only have the `trigger` field while other examples have the `trigger1` and `trigger2` field without the `trigger` or `presupposition` field.
|
| 140 |
+
Nominally, most examples look like the example in the Data Instances section above. Occassionally, however, some examples will look like:
|
| 141 |
+
```buildoutcfg
|
| 142 |
+
{
|
| 143 |
+
'sentence1': 'Did that committee know when Lissa walked through the cafe?',
|
| 144 |
+
'sentence2': 'That committee knew when Lissa walked through the cafe.',
|
| 145 |
+
'trigger1': 'interrogative',
|
| 146 |
+
'trigger2': 'unembedded',
|
| 147 |
+
'gold_label': 'neutral',
|
| 148 |
+
'control_item': True,
|
| 149 |
+
'UID': 'question_presupposition',
|
| 150 |
+
'pairID': '1821n',
|
| 151 |
+
'paradigmID': 95
|
| 152 |
+
}
|
| 153 |
+
```
|
| 154 |
+
In this example, `trigger1` and `trigger2` appear and `presupposition` and `trigger` are removed. This maintains the length of the dictionary.
|
| 155 |
+
To account for these examples, we have thus introduced the mapping above such that all examples accessed through the HF Datasets interface will have the same size as well as the same fields.
|
| 156 |
+
In the event that an example does not have a value for one of the fields, the field is maintained in the dictionary but given a value of `Not_In_Example`.
|
| 157 |
+
|
| 158 |
+
To illustrate this point, the example given in the Data Instances section above would look like the following in the HF Datasets:
|
| 159 |
+
```buildoutcfg
|
| 160 |
+
{
|
| 161 |
+
"premise": "All ten guys that proved to boast might have been divorcing.",
|
| 162 |
+
"hypothesis": "There are exactly ten guys that proved to boast.",
|
| 163 |
+
"trigger": "modal",
|
| 164 |
+
"trigger1": "Not_In_Example",
|
| 165 |
+
"trigger2": "Not_In_Example"
|
| 166 |
+
"presupposition": "positive",
|
| 167 |
+
"gold_label": "entailment",
|
| 168 |
+
"UID": "all_n_presupposition",
|
| 169 |
+
"pairID": "9e",
|
| 170 |
+
"paradigmID": 0
|
| 171 |
+
}
|
| 172 |
+
```
|
| 173 |
+
|
| 174 |
+
Below is description of the fields:
|
| 175 |
+
```buildoutcfg
|
| 176 |
+
"premise": The premise.
|
| 177 |
+
"hypothesis": The hypothesis.
|
| 178 |
+
"trigger": A detailed discussion of trigger types appears in the paper.
|
| 179 |
+
"trigger1": A detailed discussion of trigger types appears in the paper.
|
| 180 |
+
"trigger2": A detailed discussion of trigger types appears in the paper.
|
| 181 |
+
"presupposition": positive or negative.
|
| 182 |
+
"gold_label": Corresponds to entailment, contradiction, or neutral.
|
| 183 |
+
"UID": Unique id.
|
| 184 |
+
"pairID": Sentence pair ID.
|
| 185 |
+
"paradigmID": ?
|
| 186 |
+
```
|
| 187 |
+
It is not immediately clear what the difference is between `trigger`, `trigger1`, and `trigger2` is or what the `paradigmID` refers to.
|
| 188 |
+
|
| 189 |
+
**Implicature**
|
| 190 |
+
|
| 191 |
+
The `implicature` fields only have the mapping below:
|
| 192 |
+
```buildoutcfg
|
| 193 |
+
"premise" -> "sentence1"
|
| 194 |
+
"hypothesis"-> "sentence2"
|
| 195 |
+
```
|
| 196 |
+
Here is a description of the fields:
|
| 197 |
+
```buildoutcfg
|
| 198 |
+
"premise": The premise.
|
| 199 |
+
"hypothesis": The hypothesis.
|
| 200 |
+
"gold_label_log": Gold label for a logical reading of the sentence pair.
|
| 201 |
+
"gold_label_prag": Gold label for a pragmatic reading of the sentence pair.
|
| 202 |
+
"spec_relation": ?
|
| 203 |
+
"item_type": ?
|
| 204 |
+
"trigger": A detailed discussion of trigger types appears in the paper.
|
| 205 |
+
"lexemes": ?
|
| 206 |
+
```
|
| 207 |
+
|
| 208 |
+
### Data Splits
|
| 209 |
+
|
| 210 |
+
As the dataset was created to test already trained models, the only split that exists is for testing.
|
| 211 |
+
|
| 212 |
+
|
| 213 |
+
## Dataset Creation
|
| 214 |
+
|
| 215 |
+
### Curation Rationale
|
| 216 |
+
|
| 217 |
+
IMPPRES was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.
|
| 218 |
+
|
| 219 |
+
### Source Data
|
| 220 |
+
|
| 221 |
+
#### Initial Data Collection and Normalization
|
| 222 |
+
|
| 223 |
+
[More Information Needed]
|
| 224 |
+
|
| 225 |
+
#### Who are the source language producers?
|
| 226 |
+
|
| 227 |
+
[More Information Needed]
|
| 228 |
+
|
| 229 |
+
### Annotations
|
| 230 |
+
|
| 231 |
+
#### Annotation process
|
| 232 |
+
|
| 233 |
+
[More Information Needed]
|
| 234 |
+
|
| 235 |
+
#### Who are the annotators?
|
| 236 |
+
|
| 237 |
+
The annotations were generated semi-automatically.
|
| 238 |
+
|
| 239 |
+
### Personal and Sensitive Information
|
| 240 |
+
|
| 241 |
+
[More Information Needed]
|
| 242 |
+
|
| 243 |
+
## Considerations for Using the Data
|
| 244 |
+
|
| 245 |
+
### Social Impact of Dataset
|
| 246 |
+
|
| 247 |
+
[More Information Needed]
|
| 248 |
+
|
| 249 |
+
### Discussion of Biases
|
| 250 |
+
|
| 251 |
+
[More Information Needed]
|
| 252 |
+
|
| 253 |
+
### Other Known Limitations
|
| 254 |
+
|
| 255 |
+
[More Information Needed]
|
| 256 |
+
|
| 257 |
+
## Additional Information
|
| 258 |
+
|
| 259 |
+
### Dataset Curators
|
| 260 |
+
|
| 261 |
+
[More Information Needed]
|
| 262 |
+
|
| 263 |
+
### Licensing Information
|
| 264 |
+
|
| 265 |
+
IMPPRES is available under a Creative Commons Attribution-NonCommercial 4.0 International Public License ("The License"). You may not use these files except in compliance with the License. Please see the LICENSE file for more information before you use the dataset.
|
| 266 |
+
|
| 267 |
+
### Citation Information
|
| 268 |
+
|
| 269 |
+
```buildoutcfg
|
| 270 |
+
@inproceedings{jeretic-etal-2020-natural,
|
| 271 |
+
title = "Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}",
|
| 272 |
+
author = "Jereti\v{c}, Paloma and
|
| 273 |
+
Warstadt, Alex and
|
| 274 |
+
Bhooshan, Suvrat and
|
| 275 |
+
Williams, Adina",
|
| 276 |
+
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
|
| 277 |
+
month = jul,
|
| 278 |
+
year = "2020",
|
| 279 |
+
address = "Online",
|
| 280 |
+
publisher = "Association for Computational Linguistics",
|
| 281 |
+
url = "https://www.aclweb.org/anthology/2020.acl-main.768",
|
| 282 |
+
doi = "10.18653/v1/2020.acl-main.768",
|
| 283 |
+
pages = "8690--8705",
|
| 284 |
+
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.",
|
| 285 |
+
}
|
| 286 |
+
```
|
dataset_infos.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"presupposition_all_n_presupposition": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_all_n_presupposition", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"all_n_presupposition": {"name": "all_n_presupposition", "num_bytes": 458492, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 458492, "size_in_bytes": 793580}, "presupposition_both_presupposition": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_both_presupposition", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"both_presupposition": {"name": "both_presupposition", "num_bytes": 432792, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 432792, "size_in_bytes": 767880}, "presupposition_change_of_state": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_change_of_state", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"change_of_state": {"name": "change_of_state", "num_bytes": 308627, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 308627, "size_in_bytes": 643715}, "presupposition_cleft_existence": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_cleft_existence", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"cleft_existence": {"name": "cleft_existence", "num_bytes": 363238, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 363238, "size_in_bytes": 698326}, "presupposition_cleft_uniqueness": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_cleft_uniqueness", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"cleft_uniqueness": {"name": "cleft_uniqueness", "num_bytes": 388779, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 388779, "size_in_bytes": 723867}, "presupposition_only_presupposition": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_only_presupposition", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"only_presupposition": {"name": "only_presupposition", "num_bytes": 349018, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 349018, "size_in_bytes": 684106}, "presupposition_possessed_definites_existence": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_possessed_definites_existence", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"possessed_definites_existence": {"name": "possessed_definites_existence", "num_bytes": 362334, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 362334, "size_in_bytes": 697422}, "presupposition_possessed_definites_uniqueness": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_possessed_definites_uniqueness", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"possessed_definites_uniqueness": {"name": "possessed_definites_uniqueness", "num_bytes": 459403, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 459403, "size_in_bytes": 794491}, "presupposition_question_presupposition": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "trigger1": {"dtype": "string", "id": null, "_type": "Value"}, "trigger2": {"dtype": "string", "id": null, "_type": "Value"}, "presupposition": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "UID": {"dtype": "string", "id": null, "_type": "Value"}, "pairID": {"dtype": "string", "id": null, "_type": "Value"}, "paradigmID": {"dtype": "int16", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "presupposition_question_presupposition", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"question_presupposition": {"name": "question_presupposition", "num_bytes": 397227, "num_examples": 1900, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 397227, "size_in_bytes": 732315}, "implicature_connectives": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label_log": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "gold_label_prag": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "spec_relation": {"dtype": "string", "id": null, "_type": "Value"}, "item_type": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "lexemes": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "implicature_connectives", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"connectives": {"name": "connectives", "num_bytes": 221868, "num_examples": 1200, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 221868, "size_in_bytes": 556956}, "implicature_gradable_adjective": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label_log": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "gold_label_prag": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "spec_relation": {"dtype": "string", "id": null, "_type": "Value"}, "item_type": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "lexemes": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "implicature_gradable_adjective", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"gradable_adjective": {"name": "gradable_adjective", "num_bytes": 153672, "num_examples": 1200, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 153672, "size_in_bytes": 488760}, "implicature_gradable_verb": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label_log": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "gold_label_prag": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "spec_relation": {"dtype": "string", "id": null, "_type": "Value"}, "item_type": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "lexemes": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "implicature_gradable_verb", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"gradable_verb": {"name": "gradable_verb", "num_bytes": 180702, "num_examples": 1200, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 180702, "size_in_bytes": 515790}, "implicature_modals": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label_log": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "gold_label_prag": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "spec_relation": {"dtype": "string", "id": null, "_type": "Value"}, "item_type": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "lexemes": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "implicature_modals", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"modals": {"name": "modals", "num_bytes": 178560, "num_examples": 1200, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 178560, "size_in_bytes": 513648}, "implicature_numerals_10_100": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label_log": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "gold_label_prag": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "spec_relation": {"dtype": "string", "id": null, "_type": "Value"}, "item_type": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "lexemes": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "implicature_numerals_10_100", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"numerals_10_100": {"name": "numerals_10_100", "num_bytes": 208620, "num_examples": 1200, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 208620, "size_in_bytes": 543708}, "implicature_numerals_2_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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label_log": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "gold_label_prag": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "spec_relation": {"dtype": "string", "id": null, "_type": "Value"}, "item_type": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "lexemes": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "implicature_numerals_2_3", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"numerals_2_3": {"name": "numerals_2_3", "num_bytes": 188784, "num_examples": 1200, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 188784, "size_in_bytes": 523872}, "implicature_quantifiers": {"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.", "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", "homepage": "https://github.com/facebookresearch/Imppres", "license": "Creative Commons Attribution-NonCommercial 4.0 International Public License", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "gold_label_log": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "gold_label_prag": {"num_classes": 3, "names": ["entailment", "neutral", "contradiction"], "names_file": null, "id": null, "_type": "ClassLabel"}, "spec_relation": {"dtype": "string", "id": null, "_type": "Value"}, "item_type": {"dtype": "string", "id": null, "_type": "Value"}, "trigger": {"dtype": "string", "id": null, "_type": "Value"}, "lexemes": {"dtype": "string", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "builder_name": "imppres", "config_name": "implicature_quantifiers", "version": {"version_str": "1.1.0", "description": null, "major": 1, "minor": 1, "patch": 0}, "splits": {"quantifiers": {"name": "quantifiers", "num_bytes": 176814, "num_examples": 1200, "dataset_name": "imppres"}}, "download_checksums": {"https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true": {"num_bytes": 335088, "checksum": "f1f4ab03aec2248dcfbcb00b80e6099c592751cca2b542b208c6cf46f2926937"}}, "download_size": 335088, "post_processing_size": null, "dataset_size": 176814, "size_in_bytes": 511902}}
|
dummy/implicature_connectives/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c851fffc8eb600cf1ed83c908d4f7154ce3c9ab1bdabf58211fdac987cbc110f
|
| 3 |
+
size 8695
|
dummy/implicature_gradable_adjective/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c851fffc8eb600cf1ed83c908d4f7154ce3c9ab1bdabf58211fdac987cbc110f
|
| 3 |
+
size 8695
|
dummy/implicature_gradable_verb/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c851fffc8eb600cf1ed83c908d4f7154ce3c9ab1bdabf58211fdac987cbc110f
|
| 3 |
+
size 8695
|
dummy/implicature_modals/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ee9dbebced20677858681c37a9db28493f1400843c90bdcfa56ae6de99e76bd5
|
| 3 |
+
size 8695
|
dummy/implicature_numerals_10_100/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ee9dbebced20677858681c37a9db28493f1400843c90bdcfa56ae6de99e76bd5
|
| 3 |
+
size 8695
|
dummy/implicature_numerals_2_3/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c70f21e53529d6c78e166835d1e8b87a0108beb527df15c0e675583daf0b1a6e
|
| 3 |
+
size 8695
|
dummy/implicature_quantifiers/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c70f21e53529d6c78e166835d1e8b87a0108beb527df15c0e675583daf0b1a6e
|
| 3 |
+
size 8695
|
dummy/presupposition_all_n_presupposition/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9a0e7c801771b67d8c52753c9e7346490d1bc8fb68aabf504aa9ae1430c7610f
|
| 3 |
+
size 8695
|
dummy/presupposition_both_presupposition/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fb92d51a3dace5dd7428b8f649f845c168ce0549949d05d5503be459dd6d72d2
|
| 3 |
+
size 8695
|
dummy/presupposition_change_of_state/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fb92d51a3dace5dd7428b8f649f845c168ce0549949d05d5503be459dd6d72d2
|
| 3 |
+
size 8695
|
dummy/presupposition_cleft_existence/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c36895c5110f139b98ec3604940d887327b5c0709d77e3b7c307b8e34d30837a
|
| 3 |
+
size 8695
|
dummy/presupposition_cleft_uniqueness/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c36895c5110f139b98ec3604940d887327b5c0709d77e3b7c307b8e34d30837a
|
| 3 |
+
size 8695
|
dummy/presupposition_only_presupposition/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c36895c5110f139b98ec3604940d887327b5c0709d77e3b7c307b8e34d30837a
|
| 3 |
+
size 8695
|
dummy/presupposition_possessed_definites_existence/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:39c908bb81160d5dc1fd5aea2dd76d5c91df4d079dc93c23b56e7d1de9d67127
|
| 3 |
+
size 8695
|
dummy/presupposition_possessed_definites_uniqueness/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:39c908bb81160d5dc1fd5aea2dd76d5c91df4d079dc93c23b56e7d1de9d67127
|
| 3 |
+
size 8695
|
dummy/presupposition_question_presupposition/1.1.0/dummy_data.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c851fffc8eb600cf1ed83c908d4f7154ce3c9ab1bdabf58211fdac987cbc110f
|
| 3 |
+
size 8695
|
imppres.py
ADDED
|
@@ -0,0 +1,278 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""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."""
|
| 16 |
+
|
| 17 |
+
from __future__ import absolute_import, division, print_function
|
| 18 |
+
|
| 19 |
+
import json
|
| 20 |
+
import os
|
| 21 |
+
|
| 22 |
+
import datasets
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
# Find for instance the citation on arxiv or on the dataset repo/website
|
| 26 |
+
_CITATION = """\
|
| 27 |
+
@inproceedings{jeretic-etal-2020-natural,
|
| 28 |
+
title = "Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}",
|
| 29 |
+
author = "Jereti\v{c}, Paloma and
|
| 30 |
+
Warstadt, Alex and
|
| 31 |
+
Bhooshan, Suvrat and
|
| 32 |
+
Williams, Adina",
|
| 33 |
+
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
|
| 34 |
+
month = jul,
|
| 35 |
+
year = "2020",
|
| 36 |
+
address = "Online",
|
| 37 |
+
publisher = "Association for Computational Linguistics",
|
| 38 |
+
url = "https://www.aclweb.org/anthology/2020.acl-main.768",
|
| 39 |
+
doi = "10.18653/v1/2020.acl-main.768",
|
| 40 |
+
pages = "8690--8705",
|
| 41 |
+
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.",
|
| 42 |
+
}
|
| 43 |
+
"""
|
| 44 |
+
|
| 45 |
+
# You can copy an official description
|
| 46 |
+
_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."""
|
| 47 |
+
_HOMEPAGE = "https://github.com/facebookresearch/Imppres"
|
| 48 |
+
_LICENSE = "Creative Commons Attribution-NonCommercial 4.0 International Public License"
|
| 49 |
+
|
| 50 |
+
# The HuggingFace dataset library don't host the datasets but only point to the original files
|
| 51 |
+
# This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
|
| 52 |
+
_URLs = {"default": "https://github.com/facebookresearch/Imppres/blob/master/dataset/IMPPRES.zip?raw=true"}
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
class Imppres(datasets.GeneratorBasedBuilder):
|
| 56 |
+
"""Each sentence type in IMPPRES is generated according to a template that specifies the linear order of the constituents in the sentence. The constituents are sampled from a vocabulary of over 3000 lexical items annotated with grammatical features needed to ensure wellformedness. We semiautomatically generate IMPPRES using a codebase developed by Warstadt et al. (2019a) and significantly expanded for the BLiMP dataset (Warstadt et al., 2019b)."""
|
| 57 |
+
|
| 58 |
+
VERSION = datasets.Version("1.1.0")
|
| 59 |
+
|
| 60 |
+
# This is an example of a dataset with multiple configurations.
|
| 61 |
+
# If you don't want/need to define several sub-sets in your dataset,
|
| 62 |
+
# just remove the BUILDER_CONFIG_CLASS and the BUILDER_CONFIGS attributes.
|
| 63 |
+
|
| 64 |
+
# If you need to make complex sub-parts in the datasets with configurable options
|
| 65 |
+
# You can create your own builder configuration class to store attribute, inheriting from datasets.BuilderConfig
|
| 66 |
+
# BUILDER_CONFIG_CLASS = MyBuilderConfig
|
| 67 |
+
|
| 68 |
+
# You will be able to load one or the other configurations in the following list with
|
| 69 |
+
# data = datasets.load_dataset('my_dataset', 'first_domain')
|
| 70 |
+
# data = datasets.load_dataset('my_dataset', 'second_domain')
|
| 71 |
+
BUILDER_CONFIGS = [
|
| 72 |
+
datasets.BuilderConfig(
|
| 73 |
+
name="presupposition_all_n_presupposition",
|
| 74 |
+
version=VERSION,
|
| 75 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 76 |
+
),
|
| 77 |
+
datasets.BuilderConfig(
|
| 78 |
+
name="presupposition_both_presupposition",
|
| 79 |
+
version=VERSION,
|
| 80 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 81 |
+
),
|
| 82 |
+
datasets.BuilderConfig(
|
| 83 |
+
name="presupposition_change_of_state",
|
| 84 |
+
version=VERSION,
|
| 85 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 86 |
+
),
|
| 87 |
+
datasets.BuilderConfig(
|
| 88 |
+
name="presupposition_cleft_existence",
|
| 89 |
+
version=VERSION,
|
| 90 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 91 |
+
),
|
| 92 |
+
datasets.BuilderConfig(
|
| 93 |
+
name="presupposition_cleft_uniqueness",
|
| 94 |
+
version=VERSION,
|
| 95 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 96 |
+
),
|
| 97 |
+
datasets.BuilderConfig(
|
| 98 |
+
name="presupposition_only_presupposition",
|
| 99 |
+
version=VERSION,
|
| 100 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 101 |
+
),
|
| 102 |
+
datasets.BuilderConfig(
|
| 103 |
+
name="presupposition_possessed_definites_existence",
|
| 104 |
+
version=VERSION,
|
| 105 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 106 |
+
),
|
| 107 |
+
datasets.BuilderConfig(
|
| 108 |
+
name="presupposition_possessed_definites_uniqueness",
|
| 109 |
+
version=VERSION,
|
| 110 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 111 |
+
),
|
| 112 |
+
datasets.BuilderConfig(
|
| 113 |
+
name="presupposition_question_presupposition",
|
| 114 |
+
version=VERSION,
|
| 115 |
+
description="Presuppositions are facts that the speaker takes for granted when uttering a sentence.",
|
| 116 |
+
),
|
| 117 |
+
datasets.BuilderConfig(
|
| 118 |
+
name="implicature_connectives",
|
| 119 |
+
version=VERSION,
|
| 120 |
+
description="Scalar implicatures are inferences which can be drawn when one member of a memorized lexical scale is uttered.",
|
| 121 |
+
),
|
| 122 |
+
datasets.BuilderConfig(
|
| 123 |
+
name="implicature_gradable_adjective",
|
| 124 |
+
version=VERSION,
|
| 125 |
+
description="Scalar implicatures are inferences which can be drawn when one member of a memorized lexical scale is uttered.",
|
| 126 |
+
),
|
| 127 |
+
datasets.BuilderConfig(
|
| 128 |
+
name="implicature_gradable_verb",
|
| 129 |
+
version=VERSION,
|
| 130 |
+
description="Scalar implicatures are inferences which can be drawn when one member of a memorized lexical scale is uttered.",
|
| 131 |
+
),
|
| 132 |
+
datasets.BuilderConfig(
|
| 133 |
+
name="implicature_modals",
|
| 134 |
+
version=VERSION,
|
| 135 |
+
description="Scalar implicatures are inferences which can be drawn when one member of a memorized lexical scale is uttered.",
|
| 136 |
+
),
|
| 137 |
+
datasets.BuilderConfig(
|
| 138 |
+
name="implicature_numerals_10_100",
|
| 139 |
+
version=VERSION,
|
| 140 |
+
description="Scalar implicatures are inferences which can be drawn when one member of a memorized lexical scale is uttered.",
|
| 141 |
+
),
|
| 142 |
+
datasets.BuilderConfig(
|
| 143 |
+
name="implicature_numerals_2_3",
|
| 144 |
+
version=VERSION,
|
| 145 |
+
description="Scalar implicatures are inferences which can be drawn when one member of a memorized lexical scale is uttered.",
|
| 146 |
+
),
|
| 147 |
+
datasets.BuilderConfig(
|
| 148 |
+
name="implicature_quantifiers",
|
| 149 |
+
version=VERSION,
|
| 150 |
+
description="Scalar implicatures are inferences which can be drawn when one member of a memorized lexical scale is uttered.",
|
| 151 |
+
),
|
| 152 |
+
]
|
| 153 |
+
|
| 154 |
+
def _info(self):
|
| 155 |
+
if (
|
| 156 |
+
"presupposition" in self.config.name
|
| 157 |
+
): # This is the name of the configuration selected in BUILDER_CONFIGS above
|
| 158 |
+
features = datasets.Features(
|
| 159 |
+
{
|
| 160 |
+
"premise": datasets.Value("string"),
|
| 161 |
+
"hypothesis": datasets.Value("string"),
|
| 162 |
+
"trigger": datasets.Value("string"),
|
| 163 |
+
"trigger1": datasets.Value("string"),
|
| 164 |
+
"trigger2": datasets.Value("string"),
|
| 165 |
+
"presupposition": datasets.Value("string"),
|
| 166 |
+
"gold_label": datasets.ClassLabel(names=["entailment", "neutral", "contradiction"]),
|
| 167 |
+
"UID": datasets.Value("string"),
|
| 168 |
+
"pairID": datasets.Value("string"),
|
| 169 |
+
"paradigmID": datasets.Value("int16")
|
| 170 |
+
# These are the features of your dataset like images, labels ...
|
| 171 |
+
}
|
| 172 |
+
)
|
| 173 |
+
else: # This is an example to show how to have different features for "first_domain" and "second_domain"
|
| 174 |
+
features = datasets.Features(
|
| 175 |
+
{
|
| 176 |
+
"premise": datasets.Value("string"),
|
| 177 |
+
"hypothesis": datasets.Value("string"),
|
| 178 |
+
"gold_label_log": datasets.ClassLabel(names=["entailment", "neutral", "contradiction"]),
|
| 179 |
+
"gold_label_prag": datasets.ClassLabel(names=["entailment", "neutral", "contradiction"]),
|
| 180 |
+
"spec_relation": datasets.Value("string"),
|
| 181 |
+
"item_type": datasets.Value("string"),
|
| 182 |
+
"trigger": datasets.Value("string"),
|
| 183 |
+
"lexemes": datasets.Value("string"),
|
| 184 |
+
# These are the features of your dataset like images, labels ...
|
| 185 |
+
}
|
| 186 |
+
)
|
| 187 |
+
return datasets.DatasetInfo(
|
| 188 |
+
# This is the description that will appear on the datasets page.
|
| 189 |
+
description=_DESCRIPTION,
|
| 190 |
+
# This defines the different columns of the dataset and their types
|
| 191 |
+
features=features, # Here we define them above because they are different between the two configurations
|
| 192 |
+
# If there's a common (input, target) tuple from the features,
|
| 193 |
+
# specify them here. They'll be used if as_supervised=True in
|
| 194 |
+
# builder.as_dataset.
|
| 195 |
+
supervised_keys=None,
|
| 196 |
+
# Homepage of the dataset for documentation
|
| 197 |
+
homepage=_HOMEPAGE,
|
| 198 |
+
# License for the dataset if available
|
| 199 |
+
license=_LICENSE,
|
| 200 |
+
# Citation for the dataset
|
| 201 |
+
citation=_CITATION,
|
| 202 |
+
)
|
| 203 |
+
|
| 204 |
+
def _split_generators(self, dl_manager):
|
| 205 |
+
"""Returns SplitGenerators."""
|
| 206 |
+
# TODO: This method is tasked with downloading/extracting the data and defining the splits depending on the configuration
|
| 207 |
+
# If several configurations are possible (listed in BUILDER_CONFIGS), the configuration selected by the user is in self.config.name
|
| 208 |
+
|
| 209 |
+
# dl_manager is a datasets.download.DownloadManager that can be used to download and extract URLs
|
| 210 |
+
# It can accept any type or nested list/dict and will give back the same structure with the url replaced with path to local files.
|
| 211 |
+
# By default the archives will be extracted and a path to a cached folder where they are extracted is returned instead of the archive
|
| 212 |
+
my_urls = _URLs["default"]
|
| 213 |
+
base_config = self.config.name.split("_")[0]
|
| 214 |
+
secondary_config = self.config.name.split(base_config + "_")[1]
|
| 215 |
+
data_dir = os.path.join(dl_manager.download_and_extract(my_urls), "IMPPRES", base_config)
|
| 216 |
+
return [
|
| 217 |
+
datasets.SplitGenerator(
|
| 218 |
+
name=secondary_config,
|
| 219 |
+
# These kwargs will be passed to _generate_examples
|
| 220 |
+
gen_kwargs={
|
| 221 |
+
"filepath": os.path.join(data_dir, secondary_config + ".jsonl"),
|
| 222 |
+
"split": "test",
|
| 223 |
+
},
|
| 224 |
+
)
|
| 225 |
+
]
|
| 226 |
+
|
| 227 |
+
def _generate_examples(self, filepath, split):
|
| 228 |
+
""" Yields examples. """
|
| 229 |
+
# TODO: This method will receive as arguments the `gen_kwargs` defined in the previous `_split_generators` method.
|
| 230 |
+
# It is in charge of opening the given file and yielding (key, example) tuples from the dataset
|
| 231 |
+
# The key is not important, it's more here for legacy reason (legacy from tfds)
|
| 232 |
+
|
| 233 |
+
with open(filepath, encoding="utf-8") as f:
|
| 234 |
+
for id_, row in enumerate(f):
|
| 235 |
+
data = json.loads(row)
|
| 236 |
+
if "presupposition" in self.config.name:
|
| 237 |
+
# for k, v in data.items():
|
| 238 |
+
# print('{}({}): {}'.format(k, type(v), v))
|
| 239 |
+
# print('-'*55)
|
| 240 |
+
|
| 241 |
+
if "trigger1" not in list(data.keys()):
|
| 242 |
+
yield id_, {
|
| 243 |
+
"premise": data["sentence1"],
|
| 244 |
+
"hypothesis": data["sentence2"],
|
| 245 |
+
"trigger": data["trigger"],
|
| 246 |
+
"trigger1": "Not_In_Example",
|
| 247 |
+
"trigger2": "Not_In_Example",
|
| 248 |
+
"presupposition": data["presupposition"],
|
| 249 |
+
"gold_label": data["gold_label"],
|
| 250 |
+
"UID": data["UID"],
|
| 251 |
+
"pairID": data["pairID"],
|
| 252 |
+
"paradigmID": data["paradigmID"],
|
| 253 |
+
}
|
| 254 |
+
else:
|
| 255 |
+
yield id_, {
|
| 256 |
+
"premise": data["sentence1"],
|
| 257 |
+
"hypothesis": data["sentence2"],
|
| 258 |
+
"trigger": "Not_In_Example",
|
| 259 |
+
"trigger1": data["trigger1"],
|
| 260 |
+
"trigger2": data["trigger2"],
|
| 261 |
+
"presupposition": "Not_In_Example",
|
| 262 |
+
"gold_label": data["gold_label"],
|
| 263 |
+
"UID": data["UID"],
|
| 264 |
+
"pairID": data["pairID"],
|
| 265 |
+
"paradigmID": data["paradigmID"],
|
| 266 |
+
}
|
| 267 |
+
|
| 268 |
+
else:
|
| 269 |
+
yield id_, {
|
| 270 |
+
"premise": data["sentence1"],
|
| 271 |
+
"hypothesis": data["sentence2"],
|
| 272 |
+
"gold_label_log": data["gold_label_log"],
|
| 273 |
+
"gold_label_prag": data["gold_label_prag"],
|
| 274 |
+
"spec_relation": data["spec_relation"],
|
| 275 |
+
"item_type": data["item_type"],
|
| 276 |
+
"trigger": data["trigger"],
|
| 277 |
+
"lexemes": data["lexemes"],
|
| 278 |
+
}
|