CZLC/sqad_3.2_filtered

Licensing: NLP Centre Web Corpus Agreement

Introduction

This dataset is a filtered version of CZLC/SQAD_3.2 dataset.

Filtering Information

We found SQAD3.0 is computation heavy for current LLMs (due to the long contexts). To reduce the budget cost, we seek to reduce the size of the dataset, while maintaining possibly hard instances. To do so, we apply the following filtering steps:

• We remove 90% of cases when the answer is in the first 50 characters of the context (following by Stefanik et al., 2023).
• We remove 90% of qa pairs with answers in the Wikipedia title.
• We remove 90% of qa with questions with heavy overlap with the context (tf-idf > 0.9).

About Dataset

SQAD is a Czech database for question answering developed in NLP laboratory at Masaryk University. Database have been harvested from Czech Wikipedia articles by students and annotated with appropriate question, answer sentence, exact answer, question type and answer type. Part of speech tagging and lemmatization by Majka pipeline.

Citing

If you use the dataset, please cite

@inproceedings{1591218,
   author = {Sabol, Radoslav and Medveď, Marek and Horák, Aleš},
   address = {Brno},
   booktitle = {Proceedings of the Thirteenth Workshop on Recent Advances in Slavonic Natural Languages Processing, RASLAN 2019},
   editor = {Horák, Aleš and Rychlý, Pavel and Rambousek, Adam},
   keywords = {question answering; QA benchmark dataset; SQAD; Czech},
   howpublished = {tištěná verze "print"},
   language = {eng},
   location = {Brno},
   isbn = {978-80-263-1530-8},
   pages = {99-108},
   publisher = {Tribun EU},
   title = {Czech Question Answering with Extended SQAD v3.0 Benchmark Dataset},
   year = {2019}
}