POS-constrained Parallel Decoding for Non-autoregressive Generation. Paper link: https://aclanthology.org/2021.acl-long.467.pdf.

We use four datasets to evaluate POSPD, which can be obtained as follows:

• XSUM https://github.com/EdinburghNLP/XSum
• ROCStories Corpus https://cs.rochester.edu/nlp/rocstories/ We randomly sample 90k/4k stories for training/validation, and the remaining 4160 for testing. Please first get access to the origin dataset and then contact us for the splited version.
• SQuAD 1.1 https://rajpurkar.github.io/SQuAD-explorer/
• WMT14 (DE-EN) https://www.statmt.org/wmt14/translation-task.html

Setup

cd fairseq9/
pip install --editable .

Here we take training POS Tagging on wmt14 as an example.

0. Getting POS tagging data for BPE words
   To tag the BPE words, we propose a simple but effective subword-level POS tagging method for our POS predictor. which can be found in data/process_pos_data.ipynb

1. Preprossing data for fairseq
   As POS predictor needs the form of training example as <source sentence, target sentence, target POS sequnence>, we process these data repectively. Then, we put them into the same folder.

• For <source sentence, target sentence>

fairseq-preprocess
--source-lang src --target-lang tgt
--trainpref data/wmt14_data/train --validpref data/wmt14_data/valid
--testpref data/wmt14_data/test
--destdir data-bin/wmt14_data/bpe
--workers 40 --joined-dictionary

• For <target POS sequnence>

fairseq-preprocess
--source-lang tgt --target-lang tgt --only-source
--trainpref data/wmt14_data/pos/train --validpref data/wmt14_data/pos/valid --testpref data/wmt14_data/pos/test
--destdir data-bin/wmt14_data/bpe/pos
--workers 40 --joined-dictionary

Then, renaming the pos file in data/wmt14_data/pos/ with ''pos'' as prefix. For example, test.src-tgt.tgt.bin should be renamed as postest.src-tgt.tgt.bin, and dict.tgt.txt should be renamed as posdict.tgt.txt.
Finally, make sure all the datafile are in sibling directory named data/wmt14_data/bpe (remeber to place the datafile in "pos/" into the upper directory ).

2. Trainining POS Tagging

python ../fairseq_cli/trainpos.py
data-bin/wmt14_data/bpe
--arch pos_transformer_v2 --share-decoder-input-output-embed
--optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0
--lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 10000
--dropout 0.3 --weight-decay 0.0001
--criterion label_smoothed_pos_cross_entropy_v2 --label-smoothing 0.1
--max-tokens 6000
--save-dir checkpoints/wmt14_transformer_pos_N6_v2_shallow
--fp16
--clip-norm 0.1
--warmup-init-lr 1e-07
--task translation_self_pos_v2
--tensorboard-logdir logfile/wmt14_transformer_pos_N6_v2_shallow
--eval-bleu
--eval-bleu-args '{"beam": 5}'
--eval-bleu-detok moses
--eval-bleu-remove-bpe
--eval-bleu-print-samples
--skip-invalid-size-inputs-valid-test
--encoder-layers 12
--decoder-layers 1

3. Generating POS Tagging

python ../fairseq_cli/generatepos.py
data-bin/wmt14_data/bpe
--path checkpoints/wmt14_transformer_pos_N6_v2_shallow/checkpoint_best.pt
--batch-size 100
--lenpen 2
--beam 5 --remove-bpe --results-path #fill with the user dir#
--task translation_self_pos_v2

4. Processing POS Tagging for CMLM and DisCo
   Here we take CMLM as an example, you can change the corresponding file names for processing DisCo. Please run data/pos_predict_result_processing_for_cmlm.ipynb

• DisCo

0. Setup

cd disco/
pip install --editable .

1. Data Processing

python preprocess.py
--source-lang src --target-lang tgt
--trainpref ../data/wmt14_data/train --validpref ../data/wmt14_data/valid --testpref ../data/wmt14_data/test
--destdir ../data-bin/wmt14_data/bpe
--workers 60 --joined-dictionary

2. Training

python train.py ../data-bin/wmt14_data/bpe --arch disco_transformer
--criterion label_smoothed_length_cross_entropy --label-smoothing 0.1 --lr 5e-4
--warmup-init-lr 1e-7 --min-lr 1e-9 --lr-scheduler inverse_sqrt --warmup-updates 10000
--optimizer adam --adam-betas '(0.9, 0.999)' --adam-eps 1e-6 --task translation_self
--max-tokens 6000 --weight-decay 0.01 --dropout 0.3 --encoder-layers 6 --encoder-embed-dim 512
--decoder-layers 6 --decoder-embed-dim 512 --fp16 --max-source-positions 10000
--max-target-positions 10000 --seed 1
--save-dir checkpoint/wmt14_without_distill/ --dynamic-masking --ignore-eos-loss --share-all-embeddings
--skip-invalid-size-inputs-valid-test