Datasets:

Languages:
English
ArXiv:
Tags:
License:
CREPE / README.md
Seacow
update readme
8aa9eb8
metadata
license: cc-by-4.0
language:
  - en

Overview

This repository contains the dataset to the paper "Causal Reasoning of Entities and Events in Procedural Texts" and the dataset Causal Reasoning of Entities and Events in Procedural Texts (CREPE).

Files

  • data_dev_v2.json is the development set of CREPE.
  • data_test_v2.json is the test set of CREPE.

Explanations of the CREPE dataset

There are 6 columns in the dataset, namely goal, steps, event, event_answer, entity, entity_answer.

  • goal denotes the goal of a procedure.

  • steps is a list containing all steps involved in a procedure.

  • event is an event whose likelihood of happening change due to the events in steps.

  • event_answer is the ground truth likelihood change. See below for glossary.

  • entity is the entity that directly relates to the event. Its state change will have a direct impact on the likelihood of the event

  • entity_answer is the ground truth entity state change. See below for glossary.

Glossary

Label Literal Mearning
0 event/entity state is less likely to happen comparing to the previous step.
1 event/entity state is equally likely to happen comparing to the previous step.
2 event/entity state is more likely to happen comparing to the previous step.

Demonstration

goal: Clean up kitchen counter

steps:

  1. Wear rubber gloves.
  2. Get towels and wipes.
  3. Use wipes to wipe kitchen counter.
  4. Use towels to clean kitchen counter.
  5. Store the gloves.

event: The likelihood that "My skin makes contact with things I touch" after the execution of each step.

ground truth answers:

Step Number Label Literal Mearning
1 0 "less likely"
2 1 "equally likely"
3 1 "equally likely"
4 1 "equally likely"
5 2 "more likely"

entity state: hands are covered

ground truth answers:

Step Number Label Literal Mearning
5 2 "more likely"
2 1 "equally likely"
3 1 "equally likely"
4 1 "equally likely"
1 0 "less likely"