Source code for the EMNLP 2022 main conference long paper "Entropy-Based Vocabulary Substitution for Incremental Learning in Multilingual Neural Machine Translation"

In this work, we propose an entropy-based vocabulary substitution (EVS) method that just needs to walk through new language pairs for incremental learning in a large-scale multilingual data updating while remaining the size of the vocabulary.

(Core) Data Preprocessing.

Standard BPE Procedure: following https://github.com/google/sentencepiece with 64k merged BPE tokens.

EVS:

After obtaining the original vocabulary and the incremental vocabulary, you can run scripts for vocabulary substitution in three modes.

• EVS (Ours)
• frequency (choose the top-K words with the highest frequency)
• combine (Expansion)

(Optional) Model Training.

This system has been tested in the following environment.

Python version == 3.7

Pytorch version == 1.8.0

Fairseq version == 0.12.0 (pip install fairseq)

Note that it only influences the training procedure of the original and incremental model. You can choose your favorite deep learning library for model training.

We build the incremental learning procedure for Multilingual Neural Machine Translation as follows:

1. Get original multilingual translation models (or train a multilingual translation model by yourself). We will provide two MNMT models and training scripts for reproducibility.

   Data url: Permission review

   Model url: Permission review

2. Preprocessing incremental data

• Data Clean (optional, if needed)

• Get Vocabulary (follow standard BPE procedure)

• Get Vocabulary Feature (generated the incremental vocabulary with features, only for EVS). We will provide a vocabulary with features for the next stage, and you can also statisfy the feature on your own dataset.

  python scripts/get_feature.py --ov 'original_vocabulary' --iv 'incremental_vocabulary' --nv 'incremental_vocabulary_with_feature'

• EVS

  python scripts/evs.py --mode 'mode_name' --ov 'original_vocabulary' --iv 'incremental_vocabulary' --nv 'new_vocabulary'

• Data Rebuilt (if evs):

  python scripts/rebuilt.py --input 'input data' --output 'output data' --vocab 'vocabulary path'

• Data Bin (all data):

  fairseq-preprocess 
  --source-lang $SRC --target-lang $TGT \
  --trainpref $trainpref \
  --validpref $validpref \
  --testpref $testpref \
  --destdir $outfile \
  --thresholdsrc 0 --thresholdtgt 0 \
  --srcdict $vocab \
  --tgtdict $vocab \
  --workers $workers

3. Incremental Training (Joint Training) We provide all runing scripts in the folder ''run_sh'' An example:

   export lang_dict='example/langs_all,txt' (path of language sets)
   export lang_pairs='' (e.g. en-cs,en-de,en-fi,en-fr,en-hi)
   fairseq-train $DATA_PATH \
   --finetune-from-model $BASE_MODEL \
   --share-all-embeddings \
   --encoder-normalize-before --decoder-normalize-before \
   --encoder-embed-dim 1024 --encoder-ffn-embed-dim 4096 --encoder-attention-heads 16 \
   --decoder-embed-dim 1024 --decoder-ffn-embed-dim 4096 --decoder-attention-heads 16 \
   --encoder-layers 6 --decoder-layers 6 \
   --left-pad-source False --left-pad-target False \
   --arch transformer \
   --task translation_multi_simple_epoch \
   --sampling-method temperature \
   --sampling-temperature 5 \
   --lang-tok-style multilingual \
   --lang-dict $LANG_DICT \
   --lang-pairs $lang_pairs\
   --encoder-langtok src \
   --decoder-langtok \
   --optimizer adam \
   --adam-betas '(0.9, 0.98)' \
   --adam-eps 1e-9 \
   --lr 5e-4 \
   --lr-scheduler inverse_sqrt \
   --warmup-updates 4000 \
   --dropout 0.3 \
   --attention-dropout 0.1 \
   --weight-decay 0.0001 \
   --criterion label_smoothed_cross_entropy\
   --label-smoothing 0.1 \
   --max-tokens 4096 \
   --save-dir $CHECKPOINT_PATH/checkpoints/ \
   --update-freq 4 \
   --max-update 500000 \
   --seed 222 --log-format simple \
   --fp16 \
   --tensorboard-logdir $CHECKPOINT_PATH/logs/ \
   --no-progress-bar \
   --ddp-backend no_c10d

Please refer to run_sh/inference.sh & run_sh/evaluate.sh

@inproceedings{huang-etal-2022-entropy,
title = "Entropy-Based Vocabulary Substitution for Incremental Learning in Multilingual Neural Machine Translation",
author = "Huang, Kaiyu  and
      Li, Peng  and
      Ma, Jin  and
      Liu, Yang",
booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
year = "2022",
publisher = "Association for Computational Linguistics",
}