A PyTorch implementation of Controllable Paraphrase Generation with a Syntactic Exemplar

• Python 3.5
• PyTorch >= 1.0
• NLTK
• tqdm
• py-rouge
• zss

• data and tags
• evaluation (including multi-bleu, METEOR and a copy of Stanford CoreNLP)
• syntactic evaluation
• pretrained model (VGVAE+LC+WN+WPL)

run_vgvae.sh is provided as an example for training new models.

python generate.py -s PATH_TO_MODEL_PICKLE -v VOCAB_PICKLE -i SYNTACTIC_SEMANTIC_TEMPLATES -r REFERENCE_FILE -bs BEAM_SIZE

The argument -r is optional. When it is specified, the following evaluation script will be executed for reporting BLUE, ROUGE-{1,2,L}, METEOR and Syntactic TED scores.

python eval.py -i INPUT_FILE -r REFERENCE_FILE

python eval_f1_acc.py -s PATH_TO_MODEL_PICKLE -v VOCAB_PICKLE -d SYNTACTIC_EVAL_DIR

python eval_sts.py -s PATH_TO_MODEL_PICKLE -v VOCAB_PICKLE -d PATH_TO_STS

@inproceedings{mchen-controllable-19,
  author    = {Mingda Chen and Qingming Tang and Sam Wiseman and Kevin Gimpel},
  title     = {Controllable Paraphrase Generation with a Syntactic Exemplar},
  booktitle = {Proc. of {ACL}},
  year      = {2019}
}