This repo contains code for paper Fast Nearest Neighbor Machine Translation.

• SacreBLEU results for WMT

• SacreBLEU results for domain adaptation

• Python >= 3.6
• PyTorch >= 1.7.1
• faiss >= 1.5.3(pip install faiss-gpu works for me, but it is not officially released by faiss team.)
• pip install -r requirements.txt
• We modify fairseq==0.10.2 to extract features used in our paper. For details of installation and how we modify the codes, see the corresponding README file.

For each sentence-pair dataset, we do the following preprocessing steps:

1. tokenize and apply BPE
2. compute source-target alignments using fast_align
3. binarize the data using pretrained seq2seq model (by fairseq)
4. extract token representations of source/target sentences

The example scripts for preprocessing domain-adaptation/WMT data are listed below:

• For Domain Adaption dataset: thirdparty/fairseq/extract_feature_scripts/prepare-domain-adapt_with_pretrained_wmt19.sh
• For WMT14 en-fr: thirdparty/fairseq/extract_feature_scripts/prepare-wmt14en2fr_with_pretrained_wmt14.sh
• For WMT19 de-en: thirdparty/fairseq/extract_feature_scripts/prepare-wmt19en2de_with_pretrained_wmt19.sh

To find token-neighbors on source side, we do the following steps:

1. build a Datastore for each token, whose keys are the token-representations, and value are its offsets(sent_id, token_id). Note that the value of Datastore here is the offsets instead of pair of values in paper due to engineering reasons. We will use the alignments from source to target at decoding stage.
2. build faiss search index for each Datastore for approximate nearest neighbors(ANN) search
3. do KNN search using each token-representation of test dataset
4. quantize token-representations on target side. (quantization of source features have already be done at step 2)

The example scripts for find knn neighbors for domain-adaptation/WMT data are listed below:

• For Domain Adaptation Dataset: fast_knn_nmt/scripts/domain-adapt/find_knn_neighbors.sh
• For WMT14 en-fr: fast_knn_nmt/scripts/wmt-en-fr/find_knn_neighbors.sh
• For WMT19 de-en: fast_knn_nmt/scripts/wmt19-de-en/find_knn_neighbors.sh

To convert pretrained fairseq Seq2Seq ckpt to do inference, use fast_knn_nmt/custom_fairseq/train/transform_ckpt.py This script would change the task/model name of pretrained fairseq checkpoint, and adding quantizer to the model.

Note that you should change TRANSFORMER_CKPT, TRANSFORMER_CKPT and QUANTIZER_PATH to your own path.

The example scripts of inference for domain-adaptation/WMT data are listed below:

• For Domain Adaptation Dataset: See fast_knn_nmt/scripts/domain-adapt/reproduce_${domain}.sh, where domain could be it, medical, koran, law or subtitles.
• For WMT14 en-fr: fast_knn_nmt/scripts/wmt-en-fr/inference.sh
• For WMT19 de-en: fast_knn_nmt/scripts/wmt19-de-en/inference.sh

Note that you should change USER_DIR, DATA_DIR, OUT_DIR, and DETOKENIZER to your own path.

@article{meng2021fast,
  title={Fast Nearest Neighbor Machine Translation},
  author={Meng, Yuxian and Li, Xiaoya and Zheng, Xiayu and Wu, Fei and Sun, Xiaofei and Zhang, Tianwei and Li, Jiwei},
  journal={arXiv preprint arXiv:2105.14528},
  year={2021}
}

If you have any issues or questions about this repo, feel free to contact [email protected].

Apache License 2.0