raphaelmerx / mambai

Compiling parallel data for the Mambai (mgm) language

Low-Resource Machine Translation through Retrieval-Augmented LLM Prompting: A Study on the Mambai Language

Data

• Training data used for prompting compiled in mambai_parallel_eng_mgm.csv, with data coming from the Mambai Languaage Manual. The split column can be train or test
• Additional test data compiled from a native Mambai speaker compiled in test_leo.json

From the Mambai Language Manual, dictionaries have also been extracted through extract_mambai_dict.ipynb and stored in eng_mgm.json and mgm_eng.json

These dictionaries were used to mine bitext from the Mambai Language Manual through hunalign, see extract_mambai_parallel.ipynb.

Getting started

1. Setup Python requirements: python3 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
2. Extract text from Mambai Language Guide: see extract_mambai_parallel.ipynb and extract_mambai_dict.ipynb
3. Run translation experiments - see below

Translation

See notebook Mambai_prompt_for_MT.ipynb. Tweak parameters in the config dict to adjust hyper parameters.