This is an implmentation of the DRESS (Deep REinforcement Sentence Simplification) model described in Sentence Simplification with Deep Reinforcement Learning

The wikismall and wikilarge datasets can be downloaded here

8 references wikilarge test set can be downloaded here https://github.com/cocoxu/simplification/tree/master/data/turkcorpus

Copyright of the newsela dataset belongs to https://newsela.com. Please contact newsela.com to obtain the dataset https://newsela.com/data/

If you are looking for system output and don't bother to install dependencies and train a model (or run a pre-trained model), the all-system-output folder is for you.

• CUDA 7.5
• Torch
• fb-python
• pysari

Note that this model is tested using an old version of torch (available here)

CUDA_VISIBLE_DEVICES=$ID th train.lua --learnZ --useGPU \
    --model EncDecAWE \
    --attention dot \
    --seqLen 85 \
    --freqCut 4 \
    --nhid 256 \
    --nin 300 \
    --nlayers 2 \
    --dropout 0.2 \
    --lr $lr \
    --valid $valid \
    --test $test \
    --optimMethod Adam \
    --save $model \
    --train $train \
    --validout $validout --testout $testout \
    --batchSize 32 \
    --validBatchSize 32 \
    --maxEpoch 30 \
    --wordEmbedding $wembed \
    --embedOption fineTune \
    --fineTuneFactor 0 \
    | tee $log

Details see experiments/wikilarge/encdeca/train.sh. Note in newsela and wikismall datasets, you should use --freqCut 3.

If you want to generate simplifications from a pre-trained Encoder-Decoder Attention model, use the following command:

CUDA_VISIBLE_DEVICES=3 th generate_pipeline.lua \
    --modelPath $model \
    --dataPath $data.test \
    --outPathRaw $output.test \
    --oriDataPath $oridata.test \
    --oriMapPath $orimap | tee $log.test

Details see experiments/wikilarge/encdeca/generate/run_std.sh.

See details in experiments/wikilarge/dress/train_lm.sh

Create dataset scripts/get_auto_encoder_data/gen_data.sh See details in experiments/wikilarge/dress/train_auto_encoder.sh

See details in experiments/wikilarge/dress/train_dress.sh. Run a pre-trained DRESS model using this script experiments/wikilarge/dress/generate/dress/run_std.sh.

To train a lexical simplification model, you need to obtain soft word alignments in the training data, which are assigned by a pre-trained Encoder-Decoder Attention model. See details in experiments/wikilarge/dress/run_align.sh.

After you obtain the alignments, you can train a lexical simplification model using experiments/wikilarge/dress/train_lexical_simp.sh.

Lastly, you can apply the lexical simplification model with DRESS experiments/wikilarge/dress/generate/dress-ls/run_std.sh.

https://drive.google.com/open?id=0B6-YKFW-MnbOTVRMSURFbXYxNjg

Please be careful about the automatic evaluation.
You can use our released code and models to produce output for different models (i.e., EncDecA, Dress and Dress-Ls). But please make sure your evaluation settings follow the settings in our paper.

The evaluation pipeline accompanied in our code released produces single reference BLEU scores.

To be consistant with previous work, you should use 8 references wikilarge test set (availabel at https://github.com/cocoxu/simplification/tree/master/data/turkcorpus)

Therefore, to get the numbers on wikilarge, you should use scripts that support multi-bleu evalution (e.g., joshua or mtevalv13a.pl).

Checkout details for BLEU evaluation of wikilarge here

Make sure your FKGL is on corpus level.

The evaluation pipeline accompanied in our code released produces sentence-level SARI scores. You can use this simplification system (available here) to produce corpus level SARI scores.

Checkout details for SARI evaluation here

@InProceedings{D17-1063,
  author = 	"Zhang, Xingxing
		and Lapata, Mirella",
  title = 	"Sentence Simplification with Deep Reinforcement Learning",
  booktitle = 	"Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
  year = 	"2017",
  publisher = 	"Association for Computational Linguistics",
  pages = 	"595--605",
  location = 	"Copenhagen, Denmark",
  url = 	"http://aclweb.org/anthology/D17-1063"
}