# English to Kimbundu Author: Daniel Whitenack ## Data - JW300 en-kmb + `fast_align` based corpus fitering - Test sets from the Masakhane common test sets: - [test.en-kmb.en](https://github.com/masakhane-io/masakhane/tree/master/jw300_utils/test/test.en-kmb.en) - [test.en-kmb.kmb](https://github.com/masakhane-io/masakhane/tree/master/jw300_utils/test/test.en-kmb.kmb) ## Model ``` initializer: "xavier" bias_initializer: "zeros" init_gain: 1.0 embed_initializer: "xavier" embed_init_gain: 1.0 tied_embeddings: True tied_softmax: True encoder: type: "transformer" num_layers: 6 num_heads: 8 embeddings: embedding_dim: 512 scale: True dropout: 0. # typically ff_size = 4 x hidden_size hidden_size: 512 ff_size: 2048 dropout: 0.3 decoder: type: "transformer" num_layers: 6 num_heads: 8 embeddings: embedding_dim: 512 scale: True dropout: 0. # typically ff_size = 4 x hidden_size hidden_size: 512 ff_size: 2048 dropout: 0.3 ``` The pre-trained model can be found [here](https://drive.google.com/drive/folders/1-3nvjGFn1gpzpLgubOf7obhs_IJFWqj5?usp=sharing). ## Analysis Example 1: ``` Source: ( read philippians 2 : 5 - 8 . ) Reference: ( tanga filipe 2 : 5 - 8 . ) Hypothesis: ( tanga filipe 2 : 5 - 8 . ) ``` Example 2: ``` Source: today , more than ever before , there are so many things that can distract us . Reference: lelu , kuene ima iavulu i tena ku tu landukisa , m’ukulu ndenge . Hypothesis: lelu , m’ukulu , kuene ima iavulu i tena ku tu landukisa . ``` Example 3: ``` Source: the apostle paul warned about what can happen if we please ourselves first . Reference: ( 1 nzuá 2 : 16 ) o poxolo phaulu ua tu dimuna ku ima i tena kubhita se tu suua ngó o ima i tua uabhela . Hypothesis: o poxolo phaulu ua dimuna o ima i tu tena kubhanga , se tu dióndo dianga kusota o kitambuijilu kia poxolo . ``` Example 4: ``` Source: what a privilege it is to live in these last days and to be part of jehovah’s incredible organization ! Reference: tua tokala ku lembalala izuua ioso kuila , ujitu ua dikota ku tokala mu kilunga kia jihova mu ixi , mu izuua isukidila - ku ! Hypothesis: ujitu ua dikota ku kala mu izuua íii isukidila - ku , ni ku bhanga mbandu ku kilunga kia jihova mu kilunga kiê ! ``` ## Results - BLEU train: 28.57 - BLEU dev: 28.81 - BLEU test: 32.76