KACC

Datasets for the ACL 2021 Findings Paper "KACC: A Multi-task Benchmark for Knowledge Abstraction, Concretization and Completion".

Datasets

Because of the file size limit of Github, you have to unzip the KACC-L dataset first.

cd Datasets/
unzip KACC-L.zip

Mapping Files

The id-name mapping files:

• ent2name.txt

  Mappings between all instances (include entities and concepts) and their names.

• rel2name.txt

  Mappings between all relations and their names.

Raw Data

For KACC-S/M/L, each of them has a Raw folder, which contains these files:

• ent-triples.txt

  Triples (entity, relation, entity) in the entity graph.

• cpt-triples.txt

  Triples (concept, conceptual-relation, concept) in the concept graph.

• cross-triples.txt

  Cross-view triples (entity, instanceOf, concept).

• 2(3)-hop-ins(sub)-triples.txt

  2-hop or 3-hop instanceOf (ins) and subclassOf (sub) triples in each dataset.

Split Data

As our experiments are conducted on KACC-M, we provide a Split folder for KACC-M. In this folder, we provide the train/valid/test sets under the folder name of each task.

Citation

If you use our data, please cite the paper:

@inproceedings{zhou2021kacc,
  title={KACC: A Multi-task Benchmark for Knowledge Abstraction, Concretization and Completion},
  author={Zhou, Jie and Hu, Shengding and Lv, Xin and Yang, Cheng and Liu, Zhiyuan and Xu, Wei and Jiang, Jie and Li, Juanzi and Sun, Maosong},
  booktitle={Proceedings of ACL 2021: Findings},
  year={2021}
}