|
|
--- |
|
|
license: apache-2.0 |
|
|
task_categories: |
|
|
- translation |
|
|
tags: |
|
|
- image |
|
|
- text |
|
|
language: |
|
|
- es |
|
|
- zh |
|
|
- en |
|
|
- fr |
|
|
- it |
|
|
- hi |
|
|
- ko |
|
|
- ja |
|
|
- pt |
|
|
- th |
|
|
- de |
|
|
pretty_name: image translation |
|
|
size_categories: |
|
|
- 100M<n<1B |
|
|
--- |
|
|
|
|
|
# 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. |
|
|
 |
|
|
|
|
|
|
|
|
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. |
