Dataset Card for An Atlas of Machine Commonsense for If-Then Reasoning - Atomic Common Sense Dataset

Table of Contents

• Dataset Description
  • Dataset Summary
  • Supported Tasks
  • Languages
• Dataset Structure
  • Data Instances
  • Data Fields
  • Data Splits
• Dataset Creation
  • Curation Rationale
  • Source Data
  • Annotations
  • Personal and Sensitive Information
• Considerations for Using the Data
  • Social Impact of Dataset
  • Discussion of Biases
  • Other Known Limitations
• Additional Information
  • Dataset Curators
  • Licensing Information
  • Citation Information

Dataset Description

• Homepage: https://homes.cs.washington.edu/~msap/atomic/
• Repository: https://homes.cs.washington.edu/~msap/atomic/
• Paper: Maarten Sap, Ronan LeBras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith & Yejin Choi (2019). ATOMIC: An Atlas of Machine Commonsense for If-Then Reasoning. AAAI

Dataset Summary

This dataset provides the template sentences and relationships defined in the ATOMIC common sense dataset. There are three splits - train, test, and dev.

From the authors.

Disclaimer/Content warning: the events in atomic have been automatically extracted from blogs, stories and books written at various times. The events might depict violent or problematic actions, which we left in the corpus for the sake of learning the (probably negative but still important) commonsense implications associated with the events. We removed a small set of truly out-dated events, but might have missed some so please email us (msap@cs.washington.edu) if you have any concerns.

For more information, see: https://homes.cs.washington.edu/~msap/atomic/

Languages

en

Dataset Structure

Data Instances

Here is one example from the atomic dataset:

{'event': "PersonX uses PersonX's ___ to obtain", 'oEffect': [], 'oReact': ['annoyed', 'angry', 'worried'], 'oWant': [], 'prefix': ['uses', 'obtain'], 'split': 'trn', 'xAttr': [], 'xEffect': [], 'xIntent': ['to have an advantage', 'to fulfill a desire', 'to get out of trouble'], 'xNeed': [], 'xReact': ['pleased', 'smug', 'excited'], 'xWant': []}

Data Fields

Notes from the authors:

• event: just a string representation of the event.
• oEffect,oReact,oWant,xAttr,xEffect,xIntent,xNeed,xReact,xWant: annotations for each of the dimensions, stored in a json-dumped string. Note: "none" means the worker explicitly responded with the empty response, whereas [] means the worker did not annotate this dimension.
• prefix: json-dumped string that represents the prefix of content words (used to make a better trn/dev/tst split).
• split: string rep of which split the event belongs to.

Data Splits

The atomic dataset has three splits: test, train and dev of the form:

Dataset Creation

Curation Rationale

This dataset was gathered and created over to assist in common sense reasoning.

Source Data

Initial Data Collection and Normalization

See the reaserch paper and website for more detail. The dataset was created by the University of Washington using crowd sourced data

Who are the source language producers?

The Atomic authors and crowd source.

Annotations

Annotation process

Human annotations directed by forms.

Who are the annotators?

Human annotations.

Personal and Sensitive Information

Unkown, but likely none.

Considerations for Using the Data

Social Impact of Dataset

The goal for the work is to help machines understand common sense.

Discussion of Biases

Since the data is human annotators, there is likely to be baised. From the authors:

Disclaimer/Content warning: the events in atomic have been automatically extracted from blogs, stories and books written at various times. The events might depict violent or problematic actions, which we left in the corpus for the sake of learning the (probably negative but still important) commonsense implications associated with the events. We removed a small set of truly out-dated events, but might have missed some so please email us (msap@cs.washington.edu) if you have any concerns.

Other Known Limitations

While there are many relationships, the data is quite sparse. Also, each item of the dataset could be expanded into multiple sentences along the vsrious dimensions, oEffect, oRect, etc.

For example, given event: "PersonX uses PersonX's ___ to obtain" and dimension oReact: "annoyed", this could be transformed into an entry:

"PersonX uses PersonX's ___ to obtain => PersonY is annoyed"

Additional Information

Dataset Curators

The authors of Aotmic at The University of Washington

Licensing Information

The Creative Commons Attribution 4.0 International License. https://creativecommons.org/licenses/by/4.0/

Citation Information

@article{Sap2019ATOMICAA, title={ATOMIC: An Atlas of Machine Commonsense for If-Then Reasoning}, author={Maarten Sap and Ronan Le Bras and Emily Allaway and Chandra Bhagavatula and Nicholas Lourie and Hannah Rashkin and Brendan Roof and Noah A. Smith and Yejin Choi}, journal={ArXiv}, year={2019}, volume={abs/1811.00146} }