This corpus was made in collaboration with Platfòm MIT-Ayiti, who kindly allowed us to utilize their the Haitian Creole educational materials to create the MIT-Haiti Corpus (An kreyòl « Bank done MIT-Ayiti »).
In this corpus you will find parallel data (ht-en, ht-fr, ht-es), sourced from multilingual childrens stories, to evaluate Machine Translation systems and monolingual Haitian data, sourced from lesson plans and blog posts, which can be used for continued pre-training, for example.
Pou aprann plis, li pòs blog la an kreyòl: https://mit-ayiti.net/entelijans-atifisyel/.
Here, we detail how we obtained the results for OPUS-MT-EN-HT, OPUS-MT-ES-HT and OPUS-MT-FR-HT on the MIT-Ayiti test sets.
The relevant dependencies can be installed using the following command:
pip install -r requirements.txt
Our results can be reproduced by running the following script:
bash experiments/get_results.sh
./data/ contains the MIT-Ayiti corpus.
./experiments/ contains the script to generate translations (translate.py) and the script to evaluate these translations (eval.py).
./experiments/outputs/ contains the translation hypotheses, as generated by translate.py, for manual inspection.