language_creators:
- expert-generated
language:
- de
- en
- ru
multilinguality:
- translation
- multilingual
license: cc-by-4.0
configs:
- config_name: de-en
data_files:
- split: test
path: data/de-en.json
- config_name: en-de
data_files:
- split: test
path: data/en-de.json
- config_name: ru-en
data_files:
- split: test
path: data/ru-en.json
IdiomsInCtx-MT Dataset
This repository contains the IdiomsInCtx-MT dataset used in our ACL 2024 paper: The Fine-Tuning Paradox: Boosting Translation Quality Without Sacrificing LLM Abilities. See this GitHub repo for the origin of the data.
Description
The dataset consists of idiomatic expressions in context and their human-written translations. There are 1000 translations per direction. The dataset covers 2 language pairs (English-German and English-Russian) with 3 translation directions:
- English → German (
en-de) - German → English (
de-en) - Russian → English (
ru-en)
The dataset is designed to evaluate the performance of large language models and machine translation systems in handling idiomatic expressions, which can be challenging due to their non-literal meanings.
Usage
>>> dataset = load_dataset("davidstap/IdiomsInCtx-MT", "de-en") # available directions: de-en, en-de, ru-en
>>> dataset
DatasetDict({
test: Dataset({
features: ['de', 'en'],
num_rows: 1000
})
})
>>> dataset['test']['de'][0]
'Es ist mir wurst, wenn du nicht kommst.'
>>> dataset['test']['en'][0]
"I couldn't care less if you don't come."
Citation
If you use this dataset in your work, please cite our paper:
@inproceedings{stap-etal-2024-fine,
title = "The Fine-Tuning Paradox: Boosting Translation Quality Without Sacrificing {LLM} Abilities",
author = "Stap, David and
Hasler, Eva and
Byrne, Bill and
Monz, Christof and
Tran, Ke",
booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
year = "2024",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.acl-long.336",
pages = "6189--6206",
}
License
This dataset is licensed under the CC-BY-NC-4.0 License.