README.md

metadata

annotations_creators:
  - Jordan Painter, Diptesh Kanojia
language:
  - en
license:
  - cc-by-sa-4.0
multilinguality:
  - monolingual
pretty_name: 'Utilising Weak Supervision to create S3D: A Sarcasm Annotated Dataset'
size_categories:
  - 100K<n<1M
source_datasets:
  - original
task_categories:
  - text-classification

Table of Contents

• Dataset Description

Utilising Weak Supervision to Create S3D: A Sarcasm Annotated Dataset

This is the repository for the S3D dataset published at EMNLP 2022. The dataset can help build sarcasm detection models.

S3D Summary

The S3D dataset is our silver standard dataset of 100,000 tweets labelled for sarcasm using weak supervision by our BERTweet-sarcasm-combined model. These tweets can be accessed by using the Twitter API so that they can be used for other experiments. S3D contains 38879 tweets labelled as sarcastic, and 61211 tweets labelled as not being sarcastic.

Data Fields

• Tweet ID: The ID of the labelled tweet
• Label: A label to denote if a given tweet is sarcastic

Data Splits

• Train: 70,000
• Valid: 15,000
• Test: 15,000