ukuxhumana 0.0.1

Neural Machine Translation for African Languages

Ukuxhumana

"Ukuxhumana" means "Communicate" in Zulu. This project is aimed at exploring ideas for using Neural Machine Translation for low-resource languages - specifically for the official languages of South Africa.

Data

Parallel Corpuses

Our parallel corpuses are from the Autshumato project. The datasets contain data that was translated by professional translators, data that was sourced as translated file pairs from translators and data obtained from Government websites and documents

Models

Two main architectures are used throughout this project, namely Convolutional Sequence to Sequence by Gehring et. al. and Transformer by Vaswani et. al. Fairseq(-py) and Tensor2Tensor were used in modeling these techniques respectively.

Results

Results are given in BLEU.

Baseline

English -> Language

Model	Setswana	isiZulu	Northern Sotho	Xitsonga	Afrikaans
Convolutional Seq2Seq	27.77 (24.18)	0.62 (0.28)	15.35 (7.41)	36.96	16.17
Convolutional Seq2Seq (40K BPE)	23.83	1.44	4.89	34.28	21.06
Convolutional Seq2Seq (8K BPE)		2.19	15.45		26.78
Transformer (uncased)	33.53	4.55	29.23	47.37	35.26
Transformer (cased)	33.12	4.45	28.71	46.95	34.81
Transformer (40k BPE) (uncased)		4.29
Transformer (40k BPE) (cased)		4.14
Transformer (8k BPE) (uncased)
Transformer (8k BPE) (cased)

Language -> English

Model	Setswana	isiZulu	Northern Sotho	Xitsonga	Afrikaans
Convolutional Seq2Seq
Transformer (uncased)
Transformer (cased)