Machine Translation Models

Machine translation is the task of translating text from one language to another. For example, from English to Spanish. Models are based on the Transformer sequence-to-sequence architecture [nlp-machine_translation5].

An example script on how to train the model can be found here: NeMo/examples/nlp/machine_translation/enc_dec_nmt.py. The default configuration file for the model can be found at: NeMo/examples/nlp/machine_translation/conf/aayn_base.yaml.

            

            
from nemo.collections.nlp.models import MTEncDecModel

# To get the list of pre-trained models
MTEncDecModel.list_available_models()

# Download and load the a pre-trained to translate from English to Spanish
model = MTEncDecModel.from_pretrained("nmt_en_es_transformer24x6")

# Translate a sentence or list of sentences
translations = model.translate(["Hello!"], source_lang="en", target_lang="es")
        

Available Models

Supervised machine translation models require parallel corpora which comprises many examples of sentences in a source language and their corresponding translation in a target language. We use parallel data formatted as separate text files for source and target languages where sentences in corresponding files are aligned like in the table below.

It is common practice to apply data cleaning, normalization, and tokenization to the data prior to training a translation model and NeMo expects already cleaned, normalized, and tokenized data. The only data pre-processing NeMo does is subword tokenization with BPE [nlp-machine_translation4].

We recommend applying the following steps to clean, normalize, and tokenize your data. All pre-trained models released, apply these data pre-processing steps.

1. Please take a look at a detailed notebook on best practices to pre-process and clean your datasets - NeMo/tutorials/nlp/Data_Preprocessing_and_Cleaning_for_NMT.ipynb

2. Language ID filtering - This step filters out examples from your training dataset that aren’t in the correct language. For example, many datasets contain examples where source and target sentences are in the same language. You can use a pre-trained language ID classifier from fastText. Install fastText and then you can then run our script using the lid.176.bin model downloaded from the fastText website.

   Copy

   Copied!

               
   
               
   python NeMo/scripts/neural_machine_translation/filter_langs_nmt.py \
     --input-src train.en \
     --input-tgt train.es \
     --output-src train_lang_filtered.en \
     --output-tgt train_lang_filtered.es \
     --source-lang en \
     --target-lang es \
     --removed-src train_noise.en \
     --removed-tgt train_noise.es \
     --fasttext-model lid.176.bin
           

3. Length filtering - We filter out sentences from the data that are below a minimum length (1) or exceed a maximum length (250). We also filter out sentences where the ratio between source and target lengths exceeds 1.3 except for English <-> Chinese models. Moses is a statistical machine translation toolkit that contains many useful pre-processing scripts.

   Copy

   Copied!

               
   
               
   perl mosesdecoder/scripts/training/clean-corpus-n.perl -ratio 1.3 train en es train.filter 1 250
           

4. Data cleaning - While language ID filtering can sometimes help with filtering out noisy sentences that contain too many punctuations, it does not help in cases where the translations are potentially incorrect, disfluent, or incomplete. We use bicleaner a tool to identify such sentences. It trains a classifier based on many features included pre-trained language model fluency, word alignment scores from a word-alignment model like Giza++ etc. We use their available pre-trained models wherever possible and train models ourselves using their framework for remaining languages. The following script applies a pre-trained bicleaner model to the data and pick sentences that are clean with probability > 0.5.

   Copy

   Copied!

               
   
               
   awk '{print "-\t-"}' train.en \
   | paste -d "\t" - train.filter.en train.filter.es \
   | bicleaner-classify - - </path/to/bicleaner.yaml> > train.en-es.bicleaner.score
           

5. Data deduplication - We use bifixer (which uses xxHash) to hash the source and target sentences based on which we remove duplicate entries from the file. You may want to do something similar to remove training examples that are in the test dataset.

   Copy

   Copied!

               
   
               
   cat train.en-es.bicleaner.score \
     | parallel -j 25 --pipe -k -l 30000 python bifixer.py --ignore-segmentation -q - - en es \
     > train.en-es.bifixer.score
   
   awk -F awk -F "\t" '!seen[$6]++' train.en-es.bifixer.score > train.en-es.bifixer.dedup.score
           

6. Filter out data that bifixer assigns probability < 0.5 to.

   Copy

   Copied!

               
   
               
   awk -F "\t" '{ if ($5>0.5) {print $3}}' train.en-es.bifixer.dedup.score > train.cleaned.en
   awk -F "\t" '{ if ($5>0.5) {print $4}}' train.en-es.bifixer.dedup.score > train.cleaned.es
           

7. Punctuation Normalization - Punctuation, especially things like quotes can be written in different ways. It’s often useful to normalize the way they appear in text. We use the moses punctuation normalizer on all languages except Chinese.

   Copy

   Copied!

               
   
               
   perl mosesdecoder/scripts/tokenizer/normalize-punctuation.perl -l es < train.cleaned.es > train.normalized.es
   perl mosesdecoder/scripts/tokenizer/normalize-punctuation.perl -l en < train.cleaned.en > train.normalized.en
           

   For example:

   Copy

   Copied!

               
   
               
   Before - Aquí se encuentran joyerías como Tiffany`s entre negocios tradicionales suizos como la confitería Sprüngli.
   After  - Aquí se encuentran joyerías como Tiffany's entre negocios tradicionales suizos como la confitería Sprüngli.
           

8. Tokenization and word segmentation for Chinese - Naturally written text often contains punctuation markers like commas, full-stops and apostrophes that are attached to words. Tokenization by just splitting a string on spaces will result in separate token IDs for very similar items like NeMo and NeMo.. Tokenization splits punctuation from the word to create two separate tokens. In the previous example NeMo. becomes NeMo . which when split by space, results in two tokens and addresses the earlier problem.

   For example:

   Copy

   Copied!

               
   
               
   Before - Especialmente porque se enfrentará "a Mathieu (Debuchy), Yohan (Cabaye) y Adil (Rami) ", recuerda.
   After  - Especialmente porque se enfrentará " a Mathieu ( Debuchy ) , Yohan ( Cabaye ) y Adil ( Rami ) " , recuerda .
           

   We use the Moses tokenizer for all languages except Chinese.

   Copy

   Copied!

               
   
               
   perl mosesdecoder/scripts/tokenizer/tokenizer.perl -l es -no-escape < train.normalized.es > train.tokenized.es
   perl mosesdecoder/scripts/tokenizer/tokenizer.perl -l en -no-escape < train.normalized.en > train.tokenized.en
           

   For languages like Chinese where there is no explicit marker like spaces that separate words, we use Jieba to segment a string into words that are space separated.

   For example:

   Copy

   Copied!

               
   
               
   Before - 同时，卫生局认为有必要接种的其他人员，包括公共部门，卫生局将主动联络有关机构取得名单后由卫生中心安排接种。
   After  - 同时 ， 卫生局 认为 有 必要 接种 的 其他 人员 ， 包括 公共部门 ， 卫生局 将 主动 联络 有关 机构 取得 名单 后 由 卫生 中心 安排 接种 。
           

Byte-pair encoding (BPE) [nlp-machine_translation4] is a sub-word tokenization algorithm that is commonly used to reduce the large vocabulary size of datasets by splitting words into frequently occuring sub-words. Currently, Machine translation only supports the YouTokenToMe BPE tokenizer. One can set the tokenization configuration as follows:

Given BPE tokenizers, and a cleaned parallel corpus, the following steps are applied to create a TranslationDataset object.

1. Text to IDs - This performs subword tokenization with the BPE model on an input string and maps it to a sequence of tokens for the source and target text.

2. Bucketing - Sentences vary in length and when creating minibatches, we’d like sentences in them to have roughly the same length to minimize the number of <pad> tokens and to maximize computational efficiency. This step groups sentences roughly the same length into buckets.

3. Batching and padding - Creates minibatches with a maximum number of tokens specified by model.{train_ds,validation_ds,test_ds}.tokens_in_batch from buckets and pads, so they can be packed into a tensor.

Datasets can be configured as follows:

When training with DistributedDataParallel, each process has its own copy of the dataset. For large datasets, this may not always fit in CPU memory. Webdatasets circumvents this problem by efficiently iterating over tar files stored on disk. Each tar file can contain hundreds to thousands of pickle files, each containing a single minibatch.

We recommend using this method when working with datasets with > 1 million sentence pairs.

Tarred datasets can be configured as follows:

Tarred datasets can be created in two ways:

1. Using the Hydra config and training script.

   For example:

   Copy

   Copied!

               
   
               
   python examples/nlp/machine_translation/enc_dec_nmt.py \
     -cn aayn_base \
     do_training=false \
     model.preproc_out_dir=/path/to/preproc_dir \
     model.train_ds.use_tarred_dataset=true \
     model.train_ds.lines_per_dataset_fragment=1000000 \
     model.train_ds.num_batches_per_tarfile=200 \
     model.train_ds.src_file_name=train.tokenized.en \
     model.train_ds.tgt_file_name=train.tokenized.es \
     model.validation_ds.src_file_name=validation.tokenized.en \
     model.validation_ds.tgt_file_name=validation.tokenized.es \
     model.encoder_tokenizer.vocab_size=32000 \
     model.decoder_tokenizer.vocab_size=32000 \
     ~model.test_ds \
     trainer.devices=[0,1,2,3] \
     trainer.accelerator='gpu' \
     +trainer.fast_dev_run=true \
     exp_manager=null \
           

   The above script processes the parallel tokenized text files into tarred datasets that are written to /path/to/preproc_dir. Since do_training is set to False, the above script only creates tarred datasets and then exits. If do_training is set True, then one of two things happen:

   1. If no tar files are present in model.preproc_out_dir, the script first creates those files and then commences training.

   2. If tar files are already present in model.preproc_out_dir, the script starts training from the provided tar files.

2. Using a separate script without Hydra.

   Tarred datasets for parallel corpora can also be created with a script that doesn’t require specifying a configs via Hydra and just uses Python argparse.

   For example:

   Copy

   Copied!

               
   
               
   python examples/nlp/machine_translation/create_tarred_parallel_dataset.py \
     --shared_tokenizer \
     --clean \
     --bpe_dropout 0.1 \
     --src_fname train.tokenized.en \
     --tgt_fname train.tokenized.es \
     --out_dir /path/to/preproc_dir \
     --vocab_size 32000 \
     --max_seq_length 512 \
     --min_seq_length 1 \
     --tokens_in_batch 8192 \
     --lines_per_dataset_fragment 1000000 \
    --num_batches_per_tarfile 200
           

The overall model consists of an encoder, decoder, and classification head. Encoders and decoders have the following configuration options:

Our pre-trained models are optimized with Adam, with a maximum learning of 0.0004, beta of (0.9, 0.98), and inverse square root learning rate schedule from [nlp-machine_translation5]. The model.optim section sets the optimization parameters.

The following script creates tarred datasets based on the provided parallel corpus and trains a model based on the base configuration from [nlp-machine_translation5].

            

            
python examples/nlp/machine_translation/enc_dec_nmt.py \
  -cn aayn_base \
  do_training=true \
  trainer.devices=8 \
  trainer.accelerator='gpu' \
  ~trainer.max_epochs \
  +trainer.max_steps=100000 \
  +trainer.val_check_interval=1000 \
  +exp_manager.exp_dir=/path/to/store/results \
  +exp_manager.create_checkpoint_callback=True \
  +exp_manager.checkpoint_callback_params.monitor=val_sacreBLEU \
  +exp_manager.checkpoint_callback_params.mode=max \
  +exp_manager.checkpoint_callback_params.save_top_k=5 \
  model.preproc_out_dir=/path/to/preproc_dir \
  model.train_ds.use_tarred_dataset=true \
  model.train_ds.lines_per_dataset_fragment=1000000 \
  model.train_ds.num_batches_per_tarfile=200 \
  model.train_ds.src_file_name=train.tokenized.en \
  model.train_ds.tgt_file_name=train.tokenized.es \
  model.validation_ds.src_file_name=validation.tokenized.en \
  model.validation_ds.tgt_file_name=validation.tokenized.es \
  model.encoder_tokenizer.vocab_size=32000 \
  model.decoder_tokenizer.vocab_size=32000 \
  ~model.test_ds \
        

The trainer keeps track of the sacreBLEU score [nlp-machine_translation2] on the provided validation set and saves the checkpoints that have the top 5 (by default) sacreBLEU scores.

At the end of training, a .nemo file is written to the result directory which allows to run inference on a test set.

To run validation on multiple datasets, specify validation_ds.src_file_name and validation_ds.tgt_file_name with a list of file paths:

            

            
model.validation_ds.src_file_name=[/data/wmt13-en-de.src,/data/wmt14-en-de.src] \
model.validation_ds.tgt_file_name=[/data/wmt13-en-de.ref,/data/wmt14-en-de.ref] \
        

When using val_loss or val_sacreBLEU for the exp_manager.checkpoint_callback_params.monitor then the 0th indexed dataset will be used as the monitor.

To use other indexes, append the index:

            

            
exp_manager.checkpoint_callback_params.monitor=val_sacreBLEU_dl_index_1
        

Multiple test datasets work exactly the same way as validation datasets, simply replace validation_ds by test_ds in the above examples.

NMT with bottleneck encoder architecture is also supported (i.e., fixed size bottleneck), along with the training of Latent Variable Models (currently VAE, and MIM).

1. Supported learning frameworks (model.model_type):

   • NLL - Conditional cross entropy (the usual NMT loss)

   • VAE - Variational Auto-Encoder (paper)

   • MIM - Mutual Information Machine (paper)

2. Supported encoder architectures (model.encoder.arch):

   • seq2seq - the usual transformer encoder without a bottleneck

   • bridge - attention bridge bottleneck (paper)

   • perceiver - Perceiver bottleneck (paper)

Detailed description of config parameters:

• model.encoder.arch=seq2seq

  • model.encoder.hidden_steps is ignored

  • model.encoder.hidden_blocks is ignored

  • model.encoder.hidden_init_method is ignored

• model.encoder.arch=bridge

  • model.encoder.hidden_steps: input is projected to the specified fixed steps

  • model.encoder.hidden_blocks: number of encoder blocks to repeat after attention bridge projection

  • model.encoder.hidden_init_method:

    • enc_shared (default) - apply encoder to inputs, than attention bridge, followed by hidden_blocks number of the same encoder (pre and post encoders share parameters)

    • identity - apply attention bridge to inputs, followed by hidden_blocks number of the same encoder

    • enc - similar to enc_shared but the initial encoder has independent parameters

• model.encoder.arch=perceiver

  • model.encoder.hidden_steps: input is projected to the specified fixed steps

  • model.encoder.hidden_blocks: number of cross-attention + self-attention blocks to repeat after initialization block (all self-attention and cross-attention share parameters)

  • model.encoder.hidden_init_method:

    • params (default) - hidden state is initialized with learned parameters followed by cross-attention with independent parameters

    • bridge - hidden state is initialized with an attention bridge

Training requires the use of the following script (instead of enc_dec_nmt.py):

            

            
python -- examples/nlp/machine_translation/enc_dec_nmt-bottleneck.py \
      --config-path=conf \
      --config-name=aayn_bottleneck \
      ...
      model.model_type=nll \
      model.non_recon_warmup_batches=7500 \
      model.encoder.arch=perceiver \
      model.encoder.hidden_steps=32 \
      model.encoder.hidden_blocks=2 \
      model.encoder.hidden_init_method=params \
      ...
        

To generate translations on a test set and compute sacreBLEU scores, run the inference script:

            

            
python examples/nlp/machine_translation/nmt_transformer_infer.py \
  --model /path/to/model.nemo \
  --srctext test.en \
  --tgtout test.en-es.translations \
  --batch_size 128 \
  --source_lang en \
  --target_lang es
        

The --srctext file must be provided before tokenization and normalization. The resulting --tgtout file is detokenized and can be used to compute sacreBLEU scores.

            

            
cat test.en-es.translations | sacrebleu test.es
        

In practice, there are a few commonly used techniques at inference to improve translation quality. NeMo implements:

1. Model Ensembling

2. Shallow Fusion decoding with transformer language models [nlp-machine_translation1]

3. Noisy-channel re-ranking [nlp-machine_translation6]

1. Model Ensembling - Given many models trained with the same encoder and decoder tokenizer, it is possible to ensemble their predictions (by averaging probabilities at each step) to generate better translations.

\[P(y_t|y_{

NOTE: It is important to make sure that all models being ensembled are trained with the same tokenizer.

The inference script will ensemble all models provided via the –model argument as a comma separated string pointing to multiple model paths.

For example, to ensemble three models /path/to/model1.nemo, /path/to/model2.nemo, /path/to/model3.nemo, run:

            

            
python examples/nlp/machine_translation/nmt_transformer_infer.py \
  --model /path/to/model1.nemo,/path/to/model2.nemo,/path/to/model3.nemo \
  --srctext test.en \
  --tgtout test.en-es.translations \
  --batch_size 128 \
  --source_lang en \
  --target_lang es
        

2. Shallow Fusion Decoding with Transformer Language Models - Given a translation model or an ensemble ot translation models, it possible to combine the scores provided by the translation model(s) and a target-side language model.

At each decoding step, the score for a particular hypothesis on the beam is given by the weighted sum of the translation model log-probabilities and lanuage model log-probabilities.

\[\mathcal{S}(y_{1\ldots n}|x;\theta_{s \rightarrow t},\theta_{t}) = \mathcal{S}(y_{1\ldots n - 1}|x;\theta_{s \rightarrow t},\theta_{t}) + \log P(y_{n}|y_{

Lambda controls the weight assigned to the language model. For now, the only family of language models supported are transformer language models trained in NeMo.

NOTE: The transformer language model needs to be trained using the same tokenizer as the decoder tokenizer in the NMT system.

For example, to ensemble three models /path/to/model1.nemo, /path/to/model2.nemo, /path/to/model3.nemo, with shallow fusion using an LM /path/to/lm.nemo

            

            
python examples/nlp/machine_translation/nmt_transformer_infer.py \
  --model /path/to/model1.nemo,/path/to/model2.nemo,/path/to/model3.nemo \
  --lm_model /path/to/lm.nemo \
  --fusion_coef 0.05 \
  --srctext test.en \
  --tgtout test.en-es.translations \
  --batch_size 128 \
  --source_lang en \
  --target_lang es
        

3. Noisy Channel Re-ranking - Unlike ensembling and shallow fusion, noisy channel re-ranking only re-ranks the final candidates produced by beam search. It does so based on three scores

1. Forward (source to target) translation model(s) log-probabilities

2. Reverse (target to source) translation model(s) log-probabilities

3. Language Model (target) log-probabilities

\[\argmax_{i} \mathcal{S}(y_i|x) = \log P(y_i|x;\theta_{s \rightarrow t}^{ens}) + \lambda_{ncr} \big( \log P(x|y_i;\theta_{t \rightarrow s}) + \log P(y_i;\theta_{t}) \big)\]

To perform noisy-channel re-ranking, first generate a .scores file that contains log-proabilities from the forward translation model for each hypothesis on the beam.

            

            
python examples/nlp/machine_translation/nmt_transformer_infer.py \
  --model /path/to/model1.nemo,/path/to/model2.nemo,/path/to/model3.nemo \
  --lm_model /path/to/lm.nemo \
  --write_scores \
  --fusion_coef 0.05 \
  --srctext test.en \
  --tgtout test.en-es.translations \
  --batch_size 128 \
  --source_lang en \
  --target_lang es
        

This will generate a scores file test.en-es.translations.scores, which is provided as input to NeMo/examples/nlp/machine_translation/noisy_channel_reranking.py

This script also requires a reverse (target to source) translation model and a target language model.

            

            
python noisy_channel_reranking.py \
    --reverse_model=/path/to/reverse_model1.nemo,/path/to/reverse_model2.nemo \
    --language_model=/path/to/lm.nemo \
    --srctext=test.en-es.translations.scores \
    --tgtout=test-en-es.ncr.translations \
    --forward_model_coef=1.0 \
    --reverse_model_coef=0.7 \
    --target_lm_coef=0.05 \
        

Pretrained BERT encoders from either HuggingFace Transformers or Megatron-LM can be used to to train NeMo NMT models.

The library flag takes values: huggingface, megatron, and nemo.

The model_name flag is used to indicate a named model architecture. For example, we can use bert_base_cased from HuggingFace or megatron-bert-345m-cased from Megatron-LM.

The pretrained flag indicates whether or not to download the pretrained weights (pretrained=True) or instantiate the same model architecture with random weights (pretrained=False).

To use a custom model architecture from a specific library, use model_name=null and then add the custom configuration under the encoder configuration.

HuggingFace

We have provided a HuggingFace config file to use with HuggingFace encoders.

To use the config file from CLI:

            

            
--config-path=conf \
--config-name=huggingface \
        

As an example, we can configure the NeMo NMT encoder to use bert-base-cased from HuggingFace by using the huggingface config file and setting

            

            
model.encoder.pretrained=true \
model.encoder.model_name=bert-base-cased \
        

To use a custom architecture from HuggingFace we can use

            

            
+model.encoder._target_=transformers.BertConfig \
+model.encoder.hidden_size=1536 \
        

Note the + symbol is needed if we’re not adding the arguments to the YAML config file.

Megatron

We have provided a Megatron config file to use with Megatron encoders.

To use the config file from CLI:

            

            
--config-path=conf \
--config-name=megatron \
        

The checkpoint_file should be the path to Megatron-LM checkpoint:

            

            
/path/to/your/megatron/checkpoint/model_optim_rng.pt
        

In case your megatron model requires model parallelism, then checkpoint_file should point to the directory containing the standard Megatron-LM checkpoint format:

            

            
3.9b_bert_no_rng
├── mp_rank_00
│   └── model_optim_rng.pt
├── mp_rank_01
│   └── model_optim_rng.pt
├── mp_rank_02
│   └── model_optim_rng.pt
└── mp_rank_03
    └── model_optim_rng.pt
        

As an example, to train a NeMo NMT model with a 3.9B Megatron BERT encoder, we would use the following encoder configuration:

            

            
model.encoder.checkpoint_file=/path/to/megatron/checkpoint/3.9b_bert_no_rng \
model.encoder.hidden_size=2560 \
model.encoder.num_attention_heads=40 \
model.encoder.num_layers=48 \
model.encoder.max_position_embeddings=512 \
        

To train a Megatron 345M BERT, we would use

            

            
model.encoder.model_name=megatron-bert-cased \
model.encoder.checkpoint_file=/path/to/your/megatron/checkpoint/model_optim_rng.pt \
model.encoder.hidden_size=1024 \
model.encoder.num_attention_heads=16 \
model.encoder.num_layers=24 \
model.encoder.max_position_embeddings=512 \
        

If the pretrained megatron model used a custom vocab file, then set:

            

            
model.encoder_tokenizer.vocab_file=/path/to/your/megatron/vocab_file.txt
model.encoder.vocab_file=/path/to/your/megatron/vocab_file.txt
        

Use encoder.model_name=megatron_bert_uncased for uncased models with custom vocabularies and use encoder.model_name=megatron_bert_cased for cased models with custom vocabularies.

[nlp-machine_translation1]

Caglar Gulcehre, Orhan Firat, Kelvin Xu, Kyunghyun Cho, Loic Barrault, Huei-Chi Lin, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. On using monolingual corpora in neural machine translation. arXiv preprint arXiv:1503.03535, 2015.

[nlp-machine_translation2]

Matt Post. A call for clarity in reporting bleu scores. arXiv preprint arXiv:1804.08771, 2018.

[nlp-machine_translation3]

Ivan Provilkov, Dmitrii Emelianenko, and Elena Voita. Bpe-dropout: simple and effective subword regularization. arXiv preprint arXiv:1910.13267, 2019.

[nlp-machine_translation4](1,2)

Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015.

[nlp-machine_translation5](1,2,3)

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 6000–6010. 2017.

[nlp-machine_translation6]

Kyra Yee, Nathan Ng, Yann N Dauphin, and Michael Auli. Simple and effective noisy channel modeling for neural machine translation. arXiv preprint arXiv:1908.05731, 2019.