English to Isoko

Author: iroro orife

Data

- The JW300 English-Isoko dataset.

Model

Default Masakhane Transformer translation model.
Link to google drive folder with models

Analysis

The dataset requires more preprocessing to remove special characters and Scripture chapters/verse names & figures. This will make the model more generally useful outside of religious text translations. More precisely, we intend to clean up the following types of sentences:

    ( Se 1 Pita 1 : 22 . )
    [ 1 ] ( edhe - ẹme avọ 1 ) 
    ( Se Nehemaya 13 : 23 , 24 . )
    ( b ) Ẹvẹ ma sai ro le utee mai tobọ ?
    ( Se Ahwo Rom 15 : 1 , 2 . )
    ( Se Eviavia 21 : 3 - 6 . )
    ( Se 1 Jọn 5 : 14 , 15 . )

Example 1

    Source:     Still , words of apology are a strong force toward making peace .
    Reference:  Ghele na , eme unu - uwou u re fi obọ họ gaga evaọ eruo udhedhẹ .
    Hypothesis: Ghele na , eme unu - uwou yọ ẹgba ologbo nọ ma re ro ru udhedhẹ .

Example 2

    Source:     We can even ask God to ‘ create in us a pure heart . ’
    Reference:  Ma rẹ sae tubẹ yare Ọghẹnẹ re ọ ‘ kẹ omai eva efuafo . ’
    Hypothesis: Ma rẹ sae tubẹ yare Ọghẹnẹ re ọ ‘ ma omai eva efuafo . ’

Results

Tokenization	BLEU dev	BLEU test
BPE	32.58	38.05
Word-level	32.38	38.91