Data
JW300 : English-Southern Ndebele
Model Architecture
Text Preprocessing
- Remove blank/empty rows : 1856(1.78 %) samples
- Removed duplicates from source text : 6335(6.20 %) samples
- Removed duplicates from target text : 410(0.43 %) samples
- Removed all numeric-only text :39(0.04 %) samples
- Removed rows where text is fewer than orequal to 8 characters long from source text: 1653(1.73 %) samples
- Removed rows where text is fewer than orequal to 8 characters long from target text: 133(0.14 %) samples
- Removed rows where text is in test set: 1049(1.12 %) samples
BPE Tokenization
- vocab size : 4000 (superior results than 10X)
Model Config
- Details in supplied config file but used fewer transformer layers than in default notebook, with more attention heads and lower embedding size
- Trained for 75000 steps
- Took few hours on a single P100 GPU on Google colab over a two days (stopped training saved best model then reloaded that model the next day)
Results
019-11-28 13:37:38,730 Hello! This is Joey-NMT.
2019-11-28 13:38:08,636 dev bleu: 14.93 [Beam search decoding with beam size = 5 and alpha = 1.0]
2019-11-28 13:39:12,496 test bleu: 4.01 [Beam search decoding with beam size = 5 and alpha = 1.0]
Download model weights from : here .