Hugging Face's logo

Datasets:
liboaccn
/
OPUS-MIT-5M

Tasks:
Translation
Modalities:
Text
Image
Formats:
json
Languages:
Spanish
Chinese
English
Size:
1M - 10M
Tags:
image
text
Libraries:
Datasets
Dask
License:
Dataset card Data Studio Files Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Multilingual Image Translation Dataset： OPUS-MIT-5M

The OPUS-MIT-5M image translation dataset is constructed by randomly sampling 5M sentence pairs from the OPUS corpus. image/png

Figure illustrates the distribution of image-text pairs across 20 language pairs within the OPUS-MIT-5M dataset.

A key goal in creating the OPUS-MIT-5M dataset is to ensure a balanced representation across languages to enable robust multilingual image translation. We endeavor to synthesize an equal number of image-text pairs for each language pair whenever possible. However, due to variations in the availability of parallel text data within the OPUS corpus, certain language pairs, specifically TH-ZH (Thai-Chinese) and HI-ZH (Hindi-Chinese), contain a lower number of synthesized images. This imbalance reflects the underlying distribution of the source data and poses a potential challenge for evaluating model performance on low-resource language pairs.

In future work, we will explore strategies to augment these low-resource subsets and further improve the cross-linguistic generalization capabilities of the model.

Downloads last month
59