Composed by 30 students from one of the author's undergraduate classes. These sentence pairs cover topics ranging from real events (e.g., Iran's plan to attack the Saudi ambassador to the U.S.) to events/characters in movies (e.g., Batman) and purely imaginary situations, largely reflecting the pop culture as perceived by the American kids born in the early 90s. Each annotated example spans four lines: the first line contains the sentence, the second line contains the target pronoun, the third line contains the two candidate antecedents, and the fourth line contains the correct antecedent. If the target pronoun appears more than once in the sentence, its first occurrence is the one to be resolved.

definite_pronoun_resolution is configured with tfds.core.dataset_builder.BuilderConfig and has the following configurations predefined (defaults to the first one):

  • plain_text (v0.0.1) (Size: 222.12 KiB): Plain text import of the Definite Pronoun Resolution Dataset.


Plain text import of the Definite Pronoun Resolution Dataset.


  • 0.0.1 (default):
  • 1.0.0: New split API (


Split Examples
ALL 1,886
TRAIN 1,322
TEST 564


    'candidates': Sequence(Text(shape=(), dtype=tf.string)),
    'label': ClassLabel(shape=(), dtype=tf.int64, num_classes=2),
    'pronoun': Text(shape=(), dtype=tf.string),
    'sentence': Text(shape=(), dtype=tf.string),


Supervised keys (for as_supervised=True)

(u'sentence', u'label')


  title={Resolving complex cases of definite pronouns: the winograd schema challenge},
  author={Rahman, Altaf and Ng, Vincent},
  booktitle={Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning},
  organization={Association for Computational Linguistics}