Dataset Card for IMPPRES

Dataset Summary

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.

Supported Tasks and Leaderboards

Natural Language Inference.

Languages

English.

Dataset Structure

Data Instances

The data consists of 2 configurations: implicature and presupposition. Each configuration consists of several different sub-datasets:

Pressupposition

• all_n_presupposition
• change_of_state
• cleft_uniqueness
• possessed_definites_existence
• question_presupposition
• both_presupposition
• cleft_existence
• only_presupposition
• possessed_definites_uniqueness

Implicature

• connectives
• gradable_adjective
• gradable_verb
• modals
• numerals_10_100
• numerals_2_3
• quantifiers

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).

Here is an instance of the raw presupposition data from any sub-dataset:

{
"sentence1": "All ten guys that proved to boast might have been divorcing.", 
"sentence2": "There are exactly ten guys that proved to boast.", 
"trigger": "modal", 
"presupposition": "positive", 
"gold_label": "entailment", 
"UID": "all_n_presupposition", 
"pairID": "9e", 
"paradigmID": 0
}

and the raw implicature data from any sub-dataset:

{
"sentence1": "That teenager couldn't yell.", 
"sentence2": "That teenager could yell.", 
"gold_label_log": "contradiction", 
"gold_label_prag": "contradiction", 
"spec_relation": "negation", 
"item_type": "control", 
"trigger": "modal", 
"lexemes": "can - have to"
}

Data Fields

Presupposition

There is a slight mapping from the raw data fields in the presupposition sub-datasets and the fields appearing in the HuggingFace Datasets. When dealing with the HF Dataset, the following mapping of fields happens:

"premise" -> "sentence1"
"hypothesis"-> "sentence2"
"trigger" -> "trigger" or "Not_In_Example"
"trigger1" -> "trigger1" or "Not_In_Example"
"trigger2" -> "trigger2" or "Not_In_Example"
"presupposition" -> "presupposition" or "Not_In_Example"
"gold_label" -> "gold_label"
"UID" -> "UID"
"pairID" -> "pairID"
"paradigmID" -> "paradigmID"

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. 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. Nominally, most examples look like the example in the Data Instances section above. Occassionally, however, some examples will look like:

{
'sentence1': 'Did that committee know when Lissa walked through the cafe?', 
'sentence2': 'That committee knew when Lissa walked through the cafe.', 
'trigger1': 'interrogative', 
'trigger2': 'unembedded', 
'gold_label': 'neutral', 
'control_item': True, 
'UID': 'question_presupposition', 
'pairID': '1821n', 
'paradigmID': 95
}

In this example, trigger1 and trigger2 appear and presupposition and trigger are removed. This maintains the length of the dictionary. 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. 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.

To illustrate this point, the example given in the Data Instances section above would look like the following in the HF Datasets:

{
"premise": "All ten guys that proved to boast might have been divorcing.", 
"hypothesis": "There are exactly ten guys that proved to boast.", 
"trigger": "modal",
"trigger1":  "Not_In_Example",
"trigger2": "Not_In_Example"
"presupposition": "positive", 
"gold_label": "entailment", 
"UID": "all_n_presupposition", 
"pairID": "9e", 
"paradigmID": 0
}

Below is description of the fields:

"premise": The premise. 
"hypothesis": The hypothesis. 
"trigger": A detailed discussion of trigger types appears in the paper.
"trigger1":  A detailed discussion of trigger types appears in the paper.
"trigger2": A detailed discussion of trigger types appears in the paper.
"presupposition": positive or negative. 
"gold_label": Corresponds to entailment, contradiction, or neutral. 
"UID": Unique id. 
"pairID": Sentence pair ID.
"paradigmID": ?

It is not immediately clear what the difference is between trigger, trigger1, and trigger2 is or what the paradigmID refers to.

Implicature

The implicature fields only have the mapping below:

"premise" -> "sentence1"
"hypothesis"-> "sentence2"

Here is a description of the fields:

"premise": The premise. 
"hypothesis": The hypothesis. 
"gold_label_log": Gold label for a logical reading of the sentence pair.
"gold_label_prag": Gold label for a pragmatic reading of the sentence pair.
"spec_relation": ?
"item_type": ?
"trigger": A detailed discussion of trigger types appears in the paper.
"lexemes": ? 

Data Splits

As the dataset was created to test already trained models, the only split that exists is for testing.

Dataset Creation

Curation Rationale

IMPPRES was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

[More Information Needed]

Who are the annotators?

The annotations were generated semi-automatically.

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

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.

Citation Information

@inproceedings{jeretic-etal-2020-natural,
    title = "Are Natural Language Inference Models {IMPPRESsive}? {L}earning {IMPlicature} and {PRESupposition}",
    author = "Jereti\v{c}, Paloma  and
      Warstadt, Alex  and
      Bhooshan, Suvrat  and
      Williams, Adina",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.768",
    doi = "10.18653/v1/2020.acl-main.768",
    pages = "8690--8705",
    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.",
}

Contributions

Thanks to @aclifton314 for adding this dataset.