DuIE is a large-scale manually annotated dataset that can be used to evaluate architecture-based knowledge extraction algorithms.

The dataset contains more than 210,000 real-world Chinese sentences, involving more than 450,000 SPO triples (i.e., Subject-Predicate-Object triples), consisting of a pre-specified structure and 49 predicates.

All sentences in this dataset are extracted from Baidu Baike and Baidu News Search. The texts in this dataset cover various fields in real-world applications, such as news, entertainment, and user-generated content.

The dataset consists of the following data:

• 214,590 sentences, of which:
  • 172,983 sentences are used as training set;
  • 21,626 sentences are for development set;
  • 19,981 sentences are used as the test set.
• 457,866 instances, of which:
  • 363,960 instances are training set;
  • 45,558 instances are development set;
  • 48,348 instances are in the test set.

Example data: