en-ts/jw300-baseline/readme.md

English to Xitsonga

Author: Laura Martinus

Data

- The JW300 dataset.

Model

- Default Masakhane Transformer translation model.
- Link to google drive folder with model(https://drive.google.com/open?id=1onvLxPsRoem2KGnykp2vLEIe4GldDl84)

Analysis

 - TODO

Results

- BLEU dev: 35.07
- BLEU test: 44.15 

- Note: It is probably best to train this model for longer, as it timed out on Google Colab
- Note: Will likely benefit from optimising the number of BPE codes

Ran with BPE set to 40 000 instead of 4 000 and got the following:
- BLEU dev: 39.01
- BLEU test: 46.41