machine-translation

Machine Translation

This folder contains information and some scripts about the trained models for Machine translation task.

Dataset

Download and extract data from the v1.0 release of mukayese-datasets.

Dataset Format

All of the translated sentences in the dataset are formatted as follows:

{
    "sentence" : "I'm going to talk today about energy and climate.",
    "translation" : "Bugun enerji ve iklim hakkında konuşacağım."
}

Baseline Models

For Convolutional Sequence to Sequence Learning (Fairseq) and Transformer, first prepare the datasets using the provided preparation scripts, then run your experiments from fairseq.sh. For mBART50 you can use experiments in train_bart50.sh.

For example, to run the baseline model for the English-Turkish translation on WMT16, run the following command:

CUDA_VISIBLE_DEVICES=0 fairseq-train \
    data-bin/wmt16-en-tr \
    --arch transformer_iwslt_de_en --share-decoder-input-output-embed \
    --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
    --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
    --dropout 0.3 --weight-decay 0.0001 \
    --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
    --max-tokens 4096 \
    --eval-bleu \
    --eval-bleu-args '{"beam": 5, "max_len_a": 1.2, "max_len_b": 10}' \
    --eval-bleu-detok moses \
    --eval-bleu-remove-bpe \
    --eval-bleu-print-samples \
    --best-checkpoint-metric bleu --maximize-best-checkpoint-metric