MCWQ (previously termed CWQ) is a multilingual KBQA dataset grounded in and executable over Wikidata. Our dataset includes questions in four languages (Hebrew, Kannada, Chinese and English), and their associated SPARQL queries. MCWQ contains 124,187 question query pairs.

The json file of the full MCWQ dataset can be downloaded at Google Drive. The three MCD splits and a random split is stored under mcwq/splits/. The gold test set is stored as mcwq/gold_test.json.

We also provide the preprocessed files using RIR (reversible intermediate representations) in this GitHub repository under mcwq/translations/.

MCWQ's details and generation method are described in our paper mentioned below.

To replicate the results reported in the paper:

For training and evaluating T5, please refer to hf/run_t5.sh;

For training and evaluating mT5, please refer to hf/run_mt5.sh;

For evaluating zero-shot cross-lingual transfer of mT5, please refer to hf/pred_eval_mt5_zero_shot.sh.

We are working on releasing the checkpoints on HuggingFace soon.

You can find the mBERT, T5 and mT5 prediction results in the Git Repo under mcwq/results/.

Monolingual Experiments:

Zero-shot cross-lingual transfer:

If you use this dataset, please cite the following:

• Ruixiang Cui, Rahul Aralikatte, Heather Lent and Daniel Hershcovich.2022. Compositional Generalization in Multilingual Semantic Parsing over Wikidata. TACL.

@article{cui-etal-2022-compositional,
    title={Compositional Generalization in Multilingual Semantic Parsing over Wikidata},
    author={Ruixiang Cui and Rahul Aralikatte and Heather Lent and Daniel Hershcovich},
    year={2022},
    journal = "Transactions of the Association for Computational Linguistics",
    publisher = "MIT Press",
    url = "https://arxiv.org/abs/2108.03509"
}

The MCWQ dataset is based on CFQ.

For questions and usage issues, please contact [email protected] .

MCWQ is released under the CC-BY license.