mctaco

  • Description:

MC-TACO is a dataset of 13k question-answer pairs that require temporal commonsense comprehension. The dataset contains five temporal properties:

  1. duration (how long an event takes)
  2. temporal ordering (typical order of events)
  3. typical time (when an event occurs)
  4. frequency (how often an event occurs)
  5. stationarity (whether a state is maintained for a very long time or indefinitely)

We hope that this dataset can promote the future exploration of this particular class of reasoning problems.

Split Examples
'test' 9,442
'validation' 3,783
  • Feature structure:
FeaturesDict({
    'answer': Text(shape=(), dtype=string),
    'category': ClassLabel(shape=(), dtype=int64, num_classes=5),
    'label': ClassLabel(shape=(), dtype=int64, num_classes=2),
    'question': Text(shape=(), dtype=string),
    'sentence': Text(shape=(), dtype=string),
})
  • Feature documentation:
Feature Class Shape Dtype Description
FeaturesDict
answer Text string
category ClassLabel int64
label ClassLabel int64
question Text string
sentence Text string
  • Citation:
@inproceedings{ZKNR19,
    author = {Ben Zhou, Daniel Khashabi, Qiang Ning and Dan Roth},
    title = {"Going on a vacation" takes longer than "Going for a walk": A Study of Temporal Commonsense Understanding },
    booktitle = {EMNLP},
    year = {2019},
}