WikiPeople was constructed as follows:

• The Wikidata dump was downloaded and the facts concerning entities of type human were extracted.
• Then, these facts were denoised. For example, facts containing element related to image were filtered out, and facts containing element in {unknown value, no values} were removed.
• Subsequently, the subsets of elements which have at least 30 mentions were selected. And the facts related to these elements were kept. Further, each fact was parsed into a set of its role-value pairs.
• The remaining facts were randomly split into training set, validation set and test set by a percentage of 80%:10%:10%.

The statistics of WikiPeople are displayed as follows, where #Train, #Valid and #Test are the sizes of the training set, the validation set and the test set, respectively.

The training facts, validation facts and test facts stored in the files n-ary_train.json, n-ary_valid.json and n-ary_test.json, respectively, are in the same format. Each line therein is a set of ("role id": "value id/list of value ids") and the arity information in form of ("N": arity), corresponding to a fact in Wikidata about a certain person. Note that all the ids, except the ones that end with "_h" or "_t", are directly adopted from Wikidata. The two types of ids ending with "_h" or "_t" are defined by WikiPeople.

Take Line 37714 in n-ary_valid.json for example:

{
  "P166_h": "Q7186", 
  "P166_t": "Q38104", 
  "N": 5, 
  "P585": ["+1903-01-01T00:00:00Z#0#0#0#9#http://www.wikidata.org/entity/Q1985727"], 
  "P1706": ["Q41269", "Q37463"]
}

This example corresponds to the fact in Wikidata: Marie Curie received Nobel Prize in Physics in 1903, together with Henri Becquerel and Pierre Curie.

The detailed description of Line 37714 is as follows (items newly defined or introdued by WikiPeople are in italics and underlined):

Since on WikiPeople, the percentage of n-ary relational facts is low (less than 12%), a new dataset, WikiPeople-n, was derived from WikiPeople. Detailedly, based on WikiPeople, all the n-ary relational facts were kept, and some binary relational facts were randomly removed to obtain the same percentage of binary and n-ary categories as in the training set on JF17K.

The statistics of WikiPeople-n are displayed as follows:

Note that, binary relational facts from the training set, and then the validation set and the test set, were randomly removed respectively. After removing some binary relational facts from the training set, some elements (roles/values) may only exist in the validation/test set. The facts in the validation set and the test set, which contain these elements, were removed first, before conducting random removal.

We vary the percentage (0%, 50%, and 100%) of binary relational facts in WikiPeople to get three datasets, WikiPeople-0bi, WikiPeople-50bi, and WikiPeople-100bi, respectively.

The statistics of these three datasets are presented as follows:

@inproceedings{NaLP,
  title={Link prediction on n-ary relational data},
  author={Guan, Saiping and Jin, Xiaolong and Wang, Yuanzhuo and Cheng, Xueqi},
  booktitle={Proceedings of the 28th International Conference on World Wide Web (WWW'19)},
  year={2019},
  pages={583--593}
}

For any questions, please contact guansaiping@ict.ac.cn or jinxiaolong@ict.ac.cn, or open an issue.