zhiqu22 / adapnoncenter

Central Language-aware Layer

This is the codes & scripts of paper that Adapting to Non-Centered Languages for Zero-shot Multilingual Translation

There are two types of proposed method as following,

Full CLL (FCLL)

Namely, all decoder layers of Transformer have constructed CLL.

Single-Disentangled CLL (SD)

Inspired by the work of Liu et al. 2021
We remove the residual connection of FFN in a middle encoder layer to weaken the linguistic features of encoding.
To keep the balance between encoding and decoding, we empirically build CLL in the middle decoder.
Specifically, given $N$ encoder and decoder layers of the transformer, we remove the residual connection of the FFN in the encoder and replace the FFN with CLL in the decoder at $N/2+1$th layer of both networks.

Examples:
$N=5$, layer index of removing residual connection and building CLL is 3;
$N=6$, layer index of removing residual connection and building CLL is 4;

BTW, the inner size, number of layers and position of cll should be updated in the codes.

Structure

| -- result.xlsx // detailed results of our main experiments
| -- CLL  
  | -- data  
    | -- __init__.py  
    | -- language_pair_dataset_cll.py
  | -- models  
    | -- __init__.py  
    | -- transformer_cll.py  
    | -- transformer_decoder_cll.py  
    | -- transformer_encoder_cll.py  
  | -- modules  
    | -- __init__.py  
    | -- transformer_layer_cll.py  
  | -- task
    | -- __init__.py  
    | -- ......
  | -- __init__.py  
| -- SD  
| -- CLL_merged  
| -- SD_merged  
| -- directionFiles  
  | -- iwslt  
    | -- train_src.txt  
    | -- train_tgt.txt  
    | -- valid_src.txt  
    | -- ......  
  | -- square  
  | -- triangle  
  | -- opus  

Note1: 'merged' means merge all dataset of different languages together based on prior works.
We use 'merge' metric for handling data in IWSLT, Europarl, and OPUS-100.
We use multilingual translation based on different languages pairs in TED.

Note2: you need to change the hyperparameters including layer number, inner size of FNN, and url of reading direction files of merged data in the codes.

Dataset

IWSLT17

We take IWSLT17 from MMCR4NLP
We merge all dataset of different languages together.

Reminder for reproduction:
The results of IWSLT17 reported in the paper have some errors since we calculated the bleu scores by un-detokenized ref file.
But it does not influence the comparison from different models involved in this work.

OPUS-100

We merge all dataset of different languages together as Zhang et al. 2020.

Square

We take Europarl from MMCR4NLP
We merge all dataset of different languages together.

Triangle

We take Europarl from MMCR4NLP
We set random seed as 1, and use np.random.shuffle to then subsample top 200K sentences in each language.

TED talk

We follow Phillip to use top 20 languages of TED.
And, we do not use 'merge' strategy in TED.

Scripts

#preprocessing_4_tokenize:
perl mosesdecoder/scripts/tokenizer/tokenizer.perl -threads 16 -l $language >> tokenized/train/${f}

#preprocessing_4_sentencepiece:
# following fairseq doc.
echo "learning joint BPE over ${TRAIN_FILES}..."
python "$SPM_TRAIN" \
    --input=$TRAIN_FILES \
    --model_prefix=iwslt.bpe \
    --vocab_size=$BPESIZE \
    --character_coverage=1.0 \
    --model_type=bpe
python "$SPM_ENCODE" \
    --model "iwslt.bpe.model" \
    --output_format=piece \
    --inputs $filename \
    --outputs tokenized/bpe/$f

#preprocessing_4_fairseq:
fairseq-preprocess --joined-dictionary --source-lang input --target-lang output \
--trainpref iwslt/train \
--validpref iwslt/valid \
--testpref iwslt/test \
--destdir iwslt-bin/baseline

#running:
CUDA_VISIBLE_DEVICES=0,1,2,3 nohup fairseq-train iwlst-bin/baseline \
--user-dir cll_merged \
--task translation_cll_merged --seed 1 \
--source-lang input --target-lang output \
--arch transformer_cll_merged --dropout 0.3 --language_num 4 \
--ddp-backend=no_c10d \
--criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
--optimizer adam --adam-betas '(0.9,0.98)' \
--lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
--share-all-embeddings \
--max-update 100000 \
--max-tokens 4000 --weight-decay 0.0001 \
--save-dir check_iwslt/1 \
--no-epoch-checkpoints \
--no-progress-bar \
--log-interval 200 > iwslt/1.log

#generate:
fairseq-generate iwlst-bin/baseline --gen-subset test \
--user dir cll_merged \
--task translation_cll_merged \
--path check_iwslt/1/checkpoint_best.pt \
--beam 4 --remove-bpe sentencepiece > iwslt/predict/seed_1/test.txt

#post-preprocessing:
# after separating files
# employ moses(https://github.com/moses-smt/mosesdecoder) to detokenize
perl $DETOKENIZER -l $tgt < iwslt/result_$method/$src'_'$tgt/test.$tgt > iwslt/result_$method/$src'-'$tgt.detokenized.txt

# calculate results by sacrebleu (https://github.com/mjpost/sacreBLEU)
sacrebleu iwslt/ref/$src'_'$tgt/test.detokenized.$tgt --tok $tok -w 4 < iwslt/result_$method/$src'-'$tgt.detokenized.txt >> $method'_'$direction.sacrebleu

adapnoncenter's Issues

valid_src_direction.txt 、 valid_tgt_direction.txt

Why are there no files in the directionFiles/iwslt folder called valid_src_direction.txt and valid_tgt_direction.txt

scripts for training baseline model

Hey, thanks for the released code.
Is it convenient to provide scripts for the IWSLT baseline?