Dataset Card for multi-figqa

Dataset Summary

A multilingual dataset of human-written creative figurative expressions in many languages (mostly metaphors and similes). The English version (with the same format) can be found here

Languages

Languages included are Hindi, Indonesian, Javanese, Kannada, Sundanese, Swahili, and Yoruba. The language codes are respectively hi, id, kn, su, sw, and yo.

Dataset Structure

Data Instances

{
    'startphrase': the phrase,
    'ending1': one possible answer,
    'ending2': another possible answer,
    'labels': 0 if ending1 is correct else 1
}

Data Splits

All data in each language is originally intended to be used as a test set for that language.

Dataset Creation

Curation Rationale

Figurative language permeates human communication, but at the same time is relatively understudied in NLP. Datasets have been created in English to accelerate progress towards measuring and improving figurative language processing in language models (LMs). However, the use of figurative language is an expression of our cultural and societal experiences, making it difficult for these phrases to be universally applicable. We created this dataset as part of an effort to introduce more culturally relevant training data for different languages and cultures.

Source Data

Who are the source language producers?

The language producers were hired to write creative sentences in their native languages.

Additional Information

Citation Information

Please use this citation if you found this helpful:

@misc{kabra2023multilingual,
      title={Multi-lingual and Multi-cultural Figurative Language Understanding}, 
      author={Anubha Kabra and Emmy Liu and Simran Khanuja and Alham Fikri Aji and Genta Indra Winata and Samuel Cahyawijaya and Anuoluwapo Aremu and Perez Ogayo and Graham Neubig},
      year={2023},
      eprint={2305.16171},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}