Awesome
StyleTransferBibleData
The code and data here are associated with "Evaluating Prose Style Transfer with the Bible" which can be found at https://royalsocietypublishing.org/doi/full/10.1098/rsos.171920
The training, testing, and development data files are provided with the target style tags at the beginning of each source line as they were used by the seq2seq model in our paper. This may be useful information for other systems, but versions of these files without the tags will be created by following the instructions below.
All code was run in a linux virtual environment with tensorflow 1.0.1, seq2seq(https://github.com/google/seq2seq), and Moses(http://www.statmt.org/moses/) all installed and working. We also use subword-nmt (https://github.com/rsennrich/subword-nmt), but the code is replicated within this repository. Run all commands from within the Code directory unless otherwise noted.
Preparation
We need to do a few things before get into the experiments.
First, upload the modified and included experiment.py file into seq2seq/seq2seq/contrib. This will allow us to run evaluation on all checkpoints of a model.
Specify the directory that holds your moses installation:
MOSES_DIR=~/mosesdecoder/
And also specify the directory that you have installed this repository into:
REP_DIR=~/StyleTransferBibleData/
Then, combine the public->public training data files back together:
cat $REP_DIR/Data/Samples/publicVersions/Train1.tgt $REP_DIR/Data/Samples/publicVersions/Train2.tgt $REP_DIR/Data/Samples/publicVersions/Train3.tgt $REP_DIR/Data/Samples/publicVersions/Train4.tgt > $REP_DIR/Data/Samples/publicVersions/Train.tgt
cat $REP_DIR/Data/Samples/publicVersions/Train1.sourc $REP_DIR/Data/Samples/publicVersions/Train2.sourc $REP_DIR/Data/Samples/publicVersions/Train3.sourc $REP_DIR/Data/Samples/publicVersions/Train4.sourc > $REP_DIR/Data/Samples/publicVersions/Train.sourc
rm $REP_DIR/Data/Samples/publicVersions/Train1.tgt
rm $REP_DIR/Data/Samples/publicVersions/Train2.tgt
rm $REP_DIR/Data/Samples/publicVersions/Train3.tgt
rm $REP_DIR/Data/Samples/publicVersions/Train4.tgt
rm $REP_DIR/Data/Samples/publicVersions/Train1.sourc
rm $REP_DIR/Data/Samples/publicVersions/Train2.sourc
rm $REP_DIR/Data/Samples/publicVersions/Train3.sourc
rm $REP_DIR/Data/Samples/publicVersions/Train4.sourc
We need to create versions of the files which are simply tokenized using nltk and files with the BPE vocabulary applied to them which we will use for Seq2Seq. From within the Code directory:
python2 simpleTokenizer.py
mkdir ../Data/BPESamples
python massApplyBPE.py ../Data/simpleTokenSamples/ ../Data/BPESamples/ ../Data/simpleTokenBPECodes30000.txt
Next we will create versions of the sourc files which don't have the target version tag prepended. These will be used for training Moses and evaluating all results:
python removeTags.py
For Moses we also need to remove special characters from some files to prevent it from breaking during the run:
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/publicVersionsToBBE/unTaggedDev.sourc > $REP_DIR/Data/Samples/publicVersionsToBBE/unTaggedDevClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/publicVersionsToBBE/unTaggedTest.sourc > $REP_DIR/Data/Samples/publicVersionsToBBE/unTaggedTestClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/YLTToBBE/unTaggedDev.sourc > $REP_DIR/Data/Samples/YLTToBBE/unTaggedDevClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/YLTToBBE/unTaggedTest.sourc > $REP_DIR/Data/Samples/YLTToBBE/unTaggedTestClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/BBEToASV/unTaggedDev.sourc > $REP_DIR/Data/Samples/BBEToASV/unTaggedDevClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/BBEToASV/unTaggedTest.sourc > $REP_DIR/Data/Samples/BBEToASV/unTaggedTestClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/KJVToASV/unTaggedDev.sourc > $REP_DIR/Data/Samples/KJVToASV/unTaggedDevClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/KJVToASV/unTaggedTest.sourc > $REP_DIR/Data/Samples/KJVToASV/unTaggedTestClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/publicVersionsToASV/unTaggedDev.sourc > $REP_DIR/Data/Samples/publicVersionsToASV/unTaggedDevClean.sourc
$MOSES_DIR/scripts/tokenizer/escape-special-chars.perl < $REP_DIR/Data/Samples/publicVersionsToASV/unTaggedTest.sourc > $REP_DIR/Data/Samples/publicVersionsToASV/unTaggedTestClean.sourc
You should now have 4 populated directories within Data. These should be Bibles, Samples, simpleTokenSamples, and BPESamples. Each of the Samples directories should contain 6 directories: publicVersions, publicVersionsToBBE, YLTToBBE, KJVToASV, and BBEToASV. These should all contain the appropriate test, train and dev files for the dataset and processing, this includes unTagged and unTagged..Clean files in the Samples directory.
To ease usage of Moses's Experiment Management System, create a MosesData directory in each of the the Data/Samples directories. Below is an example of how to do this for the KJVToASV data
mkdir -p $REP_DIR/Data/Samples/KJVToASV/MosesData
Populate it with the correct files for the version:
cp $REP_DIR/Data/Samples/KJVToASV/unTaggedTrain.sourc $REP_DIR/Data/Samples/KJVToASV/MosesData/Train.sourc
cp $REP_DIR/Data/Samples/KJVToASV/unTaggedDevClean.sourc $REP_DIR/Data/Samples/KJVToASV/MosesData/Dev.sourc
cp $REP_DIR/Data/Samples/KJVToASV/unTaggedTestClean.sourc $REP_DIR/Data/Samples/KJVToASV/MosesData/Test.sourc
cp $REP_DIR/Data/Samples/KJVToASV/Train.tgt $REP_DIR/Data/Samples/KJVToASV/MosesData/Train.tgt
cp $REP_DIR/Data/Samples/KJVToASV/Test.tgt $REP_DIR/Data/Samples/KJVToASV/MosesData/Test.tgt
cp $REP_DIR/Data/Samples/KJVToASV/Dev.tgt $REP_DIR/Data/Samples/KJVToASV/MosesData/Dev.tgt
Now you are ready to begin the various experiments.
To train and evaluate a Seq2Seq model:
Make sure we have done all of the correct BPE data set up. This means we simply tokenize the raw samples, and apply the BPE vocab. All of these files should be in Data/BPESamples
Set names of directories to use. DATA_SET should be named the same as one of the directories in Data/BPESamples to match which pairing's data you want to use as training and development:
DATA_SET=publicVersions
MODEL_DIR=$REP_DIR/Models/${DATA_SET}Seq2Seq
mkdir -p $MODEL_DIR
Run model on the chosen data. You may want to modify train_steps and save_checkpoints_steps below for different data sets. :
python -m bin.train \
--config_paths="
$REP_DIR/Code/Seq2SeqConfigs/config.yml,
$REP_DIR/Code/Seq2SeqConfigs/trainSeq2Seq.yml,
$REP_DIR/Code/Seq2SeqConfigs/text_metrics_bpe.yml" \
--model_params "
vocab_source: $REP_DIR/Data/simpleTokenBPEVocab30000Clean.txt
vocab_target: $REP_DIR/Data/simpleTokenBPEVocab30000Clean.txt" \
--input_pipeline_train "
class: ParallelTextInputPipeline
params:
shuffle: True
source_files:
- $REP_DIR/Data/BPESamples/${DATA_SET}/Train.sourc
target_files:
- $REP_DIR/Data/BPESamples/${DATA_SET}/Train.tgt" \
--input_pipeline_dev "
class: ParallelTextInputPipeline
params:
source_files:
- $REP_DIR/Data/BPESamples/${DATA_SET}/Dev.sourc
target_files:
- $REP_DIR/Data/BPESamples/${DATA_SET}/Dev.tgt" \
--batch_size 64 \
--train_steps 200000 \
--keep_checkpoint_max 0 \
--save_checkpoints_steps 5000 \
--schedule train \
--output_dir $MODEL_DIR
Evaluate all saved model checkpoints on a sample of the dev data:
python -m bin.train \
--config_paths="
$REP_DIR/Code/Seq2SeqConfigs/config.yml,
$REP_DIR/Code/Seq2SeqConfigs/trainSeq2Seq.yml,
$REP_DIR/Code/Seq2SeqConfigs/text_metrics_bpe.yml" \
--model_params "
vocab_source: $REP_DIR/Data/simpleTokenBPEVocab30000Clean.txt
vocab_target: $REP_DIR/Data/simpleTokenBPEVocab30000Clean.txt" \
--input_pipeline_train "
class: ParallelTextInputPipeline
params:
shuffle: True
source_files:
- $REP_DIR/Data/BPESamples/${DATA_SET}/Train.sourc
target_files:
- $REP_DIR/Data/BPESamples/${DATA_SET}/Train.tgt" \
--input_pipeline_dev "
class: ParallelTextInputPipeline
params:
source_files:
- $REP_DIR/Data/BPESamples/${DATA_SET}/Dev.sourc
target_files:
- $REP_DIR/Data/BPESamples/${DATA_SET}/Dev.tgt" \
--batch_size 64 \
--schedule evalAllCheckpoints \
--output_dir $MODEL_DIR
once all checkpoints have been evaluated look at them using tensorboard.
tensorboard --logdir $MODEL_DIR
We use the loss on eval_all and identify the lowest value as our best checkpoint. If the lowest value is the final one then you should increase train_steps and continue training.
export BEST_CKPT=130002
Decode using this best checkpoint. TEST_SET should again be named the same as one of the directories in Data/BPESamples, this time to indicate which data you want to decode:
export PRED_DIR=$MODEL_DIR/decode$BEST_CKPT
export TEST_SET=YLTToBBE
mkdir -p ${PRED_DIR}
python -m bin.infer \
--tasks "
- class: DecodeText"\
--model_dir $MODEL_DIR \
--checkpoint_path ${MODEL_DIR}/model.ckpt-$BEST_CKPT \
--model_params "
inference.beam_search.beam_width: 10" \
--input_pipeline "
class: ParallelTextInputPipeline
params:
source_files:
- $REP_DIR/Data/BPESamples/${TEST_SET}/Test.sourc" \
> ${PRED_DIR}/${TEST_SET}TestPredictions.txt
This should create a file in your $PRED_DIR containing the decoded output. Since the model was trained on the files with the BPE vocab applied, change the results of decoding back to normal English:
cp ${PRED_DIR}/${TEST_SET}TestPredictions.txt ${PRED_DIR}/${TEST_SET}TestPredictionsClean.txt
sed -i 's/@@ //g' ${PRED_DIR}/${TEST_SET}TestPredictionsClean.txt
Also we will undo the NLTK tokenization that we did at the beginning.
python deTokenizeFile.py -s ${PRED_DIR}/${TEST_SET}TestPredictionsClean.txt -o ${PRED_DIR}/${TEST_SET}TestPredictionsClean.txt
Evaluate the performance using BLEU and PINC:
perl $REP_DIR/Code/multi-bleu.perl $REP_DIR/Data/Samples/${TEST_SET}/Test.tgt < ${PRED_DIR}/${TEST_SET}TestPredictionsClean.txt
perl $REP_DIR/Code/PINC.perl $REP_DIR/Data/Samples/${TEST_SET}/unTaggedTest.sourc < ${PRED_DIR}/${TEST_SET}TestPredictionsClean.txt
You can repeat the decoding, cleaning, and scoring for each test data you would like to evaluate this model on.
To train and evaluate a Moses model:
Set names of directories to use. DATA_SET should be named the same as one of the directories in Data/Samples to match which pairing's data you want to use as training and development. There should be unTagged...Clean.sourc files for the Test, Dev, and Train as created above.:
DATA_SET=KJVToASV
MODEL_DIR=$REP_DIR/Models/${DATA_SET}Moses
mkdir -p $MODEL_DIR
Including where you have moses installed, so you may need to change this
MOSES_DIR=~/mosesdecoder/
Now you will need to create a configuration file for Moses's Experiment Management System. An example is provided in Code/MosesConfigs/KJVToASVConfig.exp. In general you may need to modify the following lines:
repository-dir = /home/kcarlson/StyleTransferBibleData #update to point to REP_DIR
working-dir = $repository-dir/Models/KJVToASVMoses #update to point to MODEL_DIR
moses-src-dir = /home/kcarlson/mosesdecoder #update to point to your installation of moses
raw-stem = $bible-data/KJVToASV/MosesData/Train #change KJVToASV to the pairing you are using for training this model
raw-corpus = $bible-data/KJVToASV/MosesData/Train.$output-extension #change KJVToASV to an appropriate pairing. When the model is being trained on only a single source and single target version like KJVToASV, this will match DATA_SET. When training from multiple sources to a single target however, such as in the training of the publicToBBE model, you should use an appropriate file which only contains each target one time, such as YLTToBBE. If you don't this will cause an error in Moses.
raw-input = $bible-data/KJVToASV/MosesData/Dev.$input-extension #change KJVToASV to the pairing you are using for training and development data
raw-reference = $bible-data/KJVToASV/MosesData/Dev.$output-extension #change KJVToASV to the pairing you are using for training and development data
[EVALUATION:BBEToASV]
raw-input = $bible-data/BBEToASV/MosesData/Test.$input-extension #for each test set you wish to decode with this model you should add a new EVALUTATION:... Entry. raw-input should point to the test input file for the pairing you want to evaluate.
Now you can run the Moses experiment:
cd $MODEL_DIR
$MOSES_DIR/scripts/ems/experiment.perl -config $REP_DIR/Code/MosesConfigs/KJVToASVConfig.exp -exec
This will have produced Moses translations. We will undo our escaping of the special characters on this output and then score them. For each [EVALUATION:....] in the Config file, run the following (shown here for [EVALUATION:BBEToASV]):
$MOSES_DIR/scripts/tokenizer/deescape-special-chars.perl < $MODEL_DIR/evaluation/BBEToASV.cleaned.1 > $MODEL_DIR/evaluation/BBEToASVUnEscaped.cleaned.1
We can then find the BLEU and PINC scores for each of the tests:
perl $REP_DIR/Code/multi-bleu.perl $REP_DIR/Data/Samples/BBEToASV/Test.tgt < $MODEL_DIR/evaluation/BBEToASVUnEscaped.cleaned.1
perl $REP_DIR/Code/PINC.perl $REP_DIR/Data/Samples/BBEToASV/unTaggedTest.sourc < $MODEL_DIR/evaluation/BBEToASVUnEscaped.cleaned.1
Our Results
We have included our own results of all runs of both Seq2Seq and Moses models. A summary showing the BLEU and PINC scores can be found in Results/Results.csv.
The actual final cleaned outputs created from all of our models can be also be found in the Results Folder. The folders within Results each correspond to a model and the training data used for it. The files within these folders each correspond to a test set which was decoded using the model. For example, Results/publicToBBEMoses/YLTToBBE.txt contains the clean output created by a Moses model which was trained using all public versions as sources and only BBE as the target and which was asked to decode the test verses of Source version YLT.