sagnikrayc/snli-cf-kaushik

Dataset Card for Counterfactually Augmented SNLI

Dataset Summary

The SNLI corpus (version 1.0) is a collection of 570k human-written English sentence pairs manually labeled for balanced classification with the labels entailment, contradiction, and neutral, supporting the task of natural language inference (NLI), also known as recognizing textual entailment (RTE). In the ICLR 2020 paper Learning the Difference that Makes a Difference with Counterfactually-Augmented Data, Kaushik et. al. provided a dataset with counterfactual perturbations on the SNLI and IMDB data. This repository contains the original and counterfactual perturbations for the SNLI data, which was generated after processing the original data from here.

Languages

The language in the dataset is English as spoken by users of the website Flickr and as spoken by crowdworkers from Amazon Mechanical Turk. The BCP-47 code for English is en.

Dataset Structure

Data Instances

For each instance, there is:

• a string for the premise,
• a string for the hypothesis,
• a label: (entailment, contradiction, neutral)
• a type: this tells whether the data point is the original SNLI data point or a counterfactual perturbation.
• an idx. The ids correspond to the original id in the SNLI data. For example, if the original SNLI instance was 4626192243.jpg#3r1e, there wil be 5 data points as follows:

{
  "idx": "4626192243.jpg#3r1e-orig",
  "premise": "A man with a beard is talking on the cellphone and standing next to someone who is lying down on the street.",
  "hypothesis": "A man is prone on the street while another man stands next to him.",
  "label": "entailment",
  "type": "original"
}
{
  "idx": "4626192243.jpg#3r1e-cf-0",
  "premise": "A man with a beard is talking on the cellphone and standing next to someone who is lying down on the street.",
  "hypothesis": "A man is talking to his wife on the cellphone.",
  "label": "neutral",
  "type": "cf"
}
{
  "idx": "4626192243.jpg#3r1e-cf-1",
  "premise": "A man with a beard is talking on the cellphone and standing next to someone who is on the street.",
  "hypothesis": "A man is prone on the street while another man stands next to him.",
  "label": "neutral",
  "type": "cf"
}
{
  "idx": "4626192243.jpg#3r1e-cf-2",
  "premise": "A man with a beard is talking on the cellphone and standing next to someone who is sitting on the street.",
  "hypothesis": "A man is prone on the street while another man stands next to him.",
  "label": "contradiction",
  "_type": "cf"
}
{
  "idx": "4626192243.jpg#3r1e-cf-3",
  "premise": "A man with a beard is talking on the cellphone and standing next to someone who is lying down on the street.",
  "hypothesis": "A man is alone on the street.",
  "label": "contradiction",
  "type": "cf"
}

Data Splits

Following SNLI, this dataset also has 3 splits: train, validation, and test. The original paper says this:

RP and RH, each comprised of 3332 pairs in train, 400 in validation, and 800 in test, leading to a total of 6664 pairs in train, 800 in validation, and 1600 in test in the revised dataset.

This means for train, there are 1666 original SNLI instances, and each has 4 counterfactual perturbations (from premise and hypothesis edit), leading to a total of 1666*5 = 8330 train data points in this dataset. Similarly, validation and test has 200 and 400 original SNLI instances respectively, consequently 1000 and 2000 instances in total.

Dataset Split	Number of Instances in Split
Train	8,330
Validation	1,000
Test	2,000