davidstap / multilingual-knn-mt

Multilingual k-Nearest-Neighbor Machine Translation

This repo contains code for the paper Multilingual k-Nearest-Neighbor Machine Translation (EMNLP 2023).

Installation

Our code builds on the knn-box toolkit. You can install the toolkit and our extensions by:

conda create -n multilingual-knn python=3.8
conda activate multilingual-knn

git clone [email protected]:davidstap/multilingual-kNN-mt.git
cd multilingual-kNN-mt
pip install --editable ./
# Installing faiss with pip is not recommended, so we use conda
conda install faiss-gpu -c pytorch

1. Download pretrained models and dataset

You can prepare pretrained models and dataset by executing the following command:

cd knnbox-scripts
bash prepare_dataset_and_model.sh

It is straightforward to make changes to this script to change model and/or datasets.

2. Build kNN-MT Datastore

After preparation, the next step is to build a datastore. The following script shows how to build a datastore from Hebrew (he) into English (en), using a multilingual model that only supports Hebrew and Arabic in and out of English directions. The result is a datastore, which is stored in the data-knnds-he_en folder. To create datastores for other languages, simply change the variables.

SRC=he
TGT=en
DATA_PATH=/path/to/preprocessed/dataset
MODEL_PATH=/path/to/preprocessed/model

CUDA_VISIBLE_DEVICES=0 python knnbox-scripts/common/validate.py $DATA_PATH \
    --task translation_multi_simple_epoch \
    --langs en,he \
    --lang-pairs en-ar,en-he,ar-en,he-en \
    --seed 222 \
    --source-lang $SRC \
    --target-lang $TGT \
    --path $MODEL_PATH/best.pt \
    --model-overrides "{'eval_bleu': False, 'required_seq_len_multiple':1, 'load_alignments': False}" \
    --dataset-impl mmap \
    --valid-subset valid \
    --skip-invalid-size-inputs-valid-test \
    --max-tokens 4096 \
    --criterion label_smoothed_cross_entropy \
    --optimizer adam --adam-betas '(0.9, 0.98)' \
    --user-dir knnbox/models \
    --arch vanilla_knn_mt@transformer \
    --knn-mode build_datastore \
    --knn-datastore-path $DATA_PATH/data-knnds-${SRC}_${TGT} \
    --share-decoder-input-output-embed \
    --decoder-langtok \
    --encoder-langtok src \
    --fp16 \
    --dropout 0.1 \
    --label-smoothing 0.1

Bilingual datastores can be combined, resulting in a multilingual datastores. This leads to better performance. The following script can be used to combine he-en and ar-en:

DATA_PATH=/path/to/preprocessed/dataset

python knnbox-scripts/common/combine_datastores.py \
    --path $DATA_PATH \
    --pairs ar_en he_en \
    --save_path $DATA_PATH/data-knnds-ar_en-he_en \

3. Do kNN-MT inference.

Use the following script to do inference using kNN, and calculate BLEU scores:

SRC=he
TGT=en
KNN_K=16
KNN_LAMBDA=0.5
KNN_TEMP=100
DATA_PATH=/path/to/preprocessed/dataset
MODEL_PATH=/path/to/preprocessed/model
RESULTS_PATH=${MODEL_PATH}/generations/gen-test-${SRC}_${TGT}-${DS}-k_${KNN_K}-l_${KNN_LAMBDA}-t_${KNN_TEMP}

CUDA_VISIBLE_DEVICES=0 python knnbox-scripts/common/generate.py $DATA_PATH \
    --task translation_multi_simple_epoch \
    --langs en,he \
    --lang-pairs en-ar,en-he,ar-en,he-en \
    --source-lang $SRC \
    --target-lang $TGT \
    --remove-bpe 'sentencepiece' \
    --path $MODEL_PATH/best.pt \
    --share-decoder-input-output-embed \
    --results-path $RESULTS_PATH \
    --dataset-impl mmap \
    --beam 5 \
    --gen-subset test \
    --max-tokens 1024 \
    --user-dir knnbox/models \
    --arch vanilla_knn_mt@transformer \
    --knn-mode inference \
    --knn-datastore-path $DATA_PATH/data-knnds-${SRC}_${TGT} \
    --knn-k $KNN_K \
    --knn-lambda $KNN_LAMBDA \
    --knn-temperature $KNN_TEMP \
    --model-overrides "{'eval_bleu': False, 'required_seq_len_multiple':1, 'load_alignments': False}" \
    --decoder-langtok \
    --encoder-langtok src \
    --skip-invalid-size-inputs-valid-test \
    --sampling-method temperature \
    --label-smoothing 0.1 \
    --criterion label_smoothed_cross_entropy \
    --fp16 \
    --seed 222

grep ^T $RESULTS_PATH/generate-test.txt | LC_ALL=C sort -V | cut -f2- > $RESULTS_PATH/ref.txt
grep ^D $RESULTS_PATH/generate-test.txt | LC_ALL=C sort -V | cut -f3- > $RESULTS_PATH/hyp.txt
grep ^S $RESULTS_PATH/generate-test.txt | LC_ALL=C sort -V | cut -f2- > $RESULTS_PATH/src.txt

echo knn inference result for $SRC-$TGT K=$KNN_K L=$KNN_LAMBDA T=$KNN_TEMP saved at $RESULTS_PATH
sacrebleu $RESULTS_PATH/ref.txt -i $RESULTS_PATH/hyp.txt -m bleu | grep '"score"' | grep -oE '[0-9]+(\.[0-9]+)?' > $RESULTS_PATH/bleu.txt
cat $RESULTS_PATH/bleu.txt

• KNN_K is the number of neighbors, good values for the TED dataset are 8, 16, 32, 64
• KNN_LAMBDA is the weight of the kNN distribution, good values are {0.2, 0.3, ..., 0.7}
• KNN_TEMP is the temperature for the kNN distribution to make it more smooth. Good values are 10, 100.

Instead of fixing these values, one can do a hyperparameter search on the validation set, and use the best parameters to do decoding on the test set.

Citation

If you use our code, please cite the following:

@misc{stap2023multilingualknn,
      title={Multilingual k-Nearest-Neighbor Machine Translation}, 
      author={David Stap and Christof Monz},
      booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
      year={2023},
      publisher = "Association for Computational Linguistics",
}