A carefully curated dataset of cryptic crosswords ⬛⬜🕵️
https://arxiv.org/pdf/2103.01242.pdf
Download the official dataset here.
For recreating the data yourself, see the scraping README. If you encounter any issues, feel free to open a GitHub issue.
https://huggingface.co/datasets/viewer/?dataset=cryptonite
!pip install datasets
from datasets import load_dataset
dataset = load_dataset('cryptonite')Current NLP datasets targeting ambiguity can be solved by a native speaker with relative ease. We present Cryptonite, a large-scale dataset based on cryptic crosswords, which is both linguistically complex and naturally sourced. Each example in Cryptonite is a cryptic clue, a short phrase or sentence with a misleading surface reading, whose solving requires disambiguating semantic, syntactic, and phonetic wordplays, as well as world knowledge. Cryptic clues pose a challenge even for experienced solvers, though top-tier experts can solve them with almost 100% accuracy. Cryptonite is a challenging task for current models; fine-tuning T5-Large on 470k cryptic clues achieves only 7.6% accuracy, on par with the accuracy of a rule-based clue solver (8.6%).
Choose your favorite citation format from our page on ACL Anthology.
Cryptonite: How I Stopped Worrying and Learned(?) to Love Ambiguity ⬛⬜🕵️