|
--- |
|
license: mit |
|
task_categories: |
|
- translation |
|
language: |
|
- fr |
|
- en |
|
size_categories: |
|
- 100K<n<1M |
|
--- |
|
|
|
# Anki-Global Voices English-French Translation Dataset |
|
|
|
## Description |
|
|
|
The Anki-Global Voices English-French Translation Dataset is a comprehensive collection of over 500,000 translation pairs, merging the Anki English to French dataset with the Global Voices English to French dataset. This unique dataset provides a wide range of sentences, suitable for training and evaluating machine translation models in both informal and formal language contexts. |
|
|
|
### Languages |
|
|
|
This dataset includes sentence pairs in English (ISO 639-1: EN) and French (ISO 639-1: FR). |
|
|
|
## Dataset Structure |
|
|
|
### Data Instances |
|
|
|
Each data instance consists of an English sentence and its French translation. Example: |
|
|
|
```json |
|
{ |
|
"source": "He is a good boy.", |
|
"target": "C'est un bon garçon." |
|
} |
|
``` |
|
|
|
### Data Fields |
|
|
|
- `source`: The English sentence. |
|
- `target`: The corresponding French translation. |
|
|
|
### Data Splits |
|
|
|
The dataset is divided into training, validation, and test sets. The distribution is as follows (illustrative numbers): |
|
|
|
- Train: 439,000 pairs |
|
- Validation: 54,900 pairs |
|
- Test: 54,900 pairs |
|
|
|
## Dataset Creation |
|
|
|
### Curation Rationale |
|
|
|
This dataset aims to aid the development of machine translation models by providing a rich source of linguistic variations and expressions in English and French, covering both everyday language and formal contexts. |
|
|
|
### Source Data |
|
|
|
#### Initial Data Collection and Normalization |
|
|
|
The dataset combines two sources: |
|
|
|
1. **Anki English to French Dataset**: Comprised of curated translation pairs submitted to the Tatoeba Project, mostly short and with casual language. [More information](https://www.manythings.org/anki/) |
|
2. **Global Voices English to French Dataset**: Derived from news and cultural articles on the Global Voices websites, featuring more formal, complicated language. [More information](https://data.europa.eu/data/datasets/elrc_2006?locale=en) |
|
|
|
Approximately 21,000 pairs were removed from the Global Voices dataset before merging with the Anki Dataset due to severe misalignment. Samples have already been shuffled. |
|
|
|
### Annotations |
|
|
|
The dataset contains translations but no additional annotations. |
|
|
|
#### Annotation Process |
|
|
|
Translations in the Anki dataset were community-contributed, and those in the Global Voices dataset were done by professional translators and journalists. |
|
|
|
### Personal and Sensitive Information |
|
|
|
The dataset does not contain personal or sensitive information. |
|
|
|
## Considerations for Using the Data |
|
|
|
### Social Impact of Dataset |
|
|
|
This dataset can help improve the accuracy of machine translation systems, facilitating communication across different language speakers. |
|
|
|
### Discussion of Biases |
|
|
|
The dataset covers both informal and formal language contexts, providing a balanced linguistic range. However, source datasets' biases might be present. |
|
|
|
### Other Known Limitations |
|
|
|
The dataset focuses exclusively on English-French translations, which may not generalize to other language pairs or cultural contexts. |
|
|
|
## Additional Information |
|
|
|
### Dataset Curators |
|
|
|
Curated by Ariel Guerra-Adames, this dataset combines the Anki and Global Voices datasets into a comprehensive resource for machine translation. |
|
|
|
### Licensing Information |
|
|
|
See LICENSE |
|
|
