This is the rewriter described in our paper:
@inproceedings{N18-1007, title = "Improved Lexically Constrained Decoding for Translation and Monolingual Rewriting", author = "Hu, J. Edward and Khayrallah, Huda and Culkin, Ryan and Xia, Patrick and Chen, Tongfei and Post, Matt and Van Durme, Benjamin", booktitle = "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)", month = jun, year = "2019", address = "Minneapolis, Minnesota", publisher = "Association for Computational Linguistics",
It is available from
https://github.com/decompositional-semantics-initiative/improved-ParaBank-rewriter/releases
You can also interact with it using our online demo.
After downloading the release, unpacking it, and changing to that directory, run the following command:
pip3 install -r requirements.txt
You may also have to install spacy models:
python3 -m spacy download en
python3 -m spacy download en_core_web_lg
The rewriter takes raw, unprocessed input and returns the same.
It applies all the pre-processing automatically, and undoes it afterwards.
To run the pipeline, use the paraphrase.sh script:
Usage:
cat sentences.txt | /path/to/rewriter/paraphrase.sh > paraphrases.txt
It uses Sockeye internally, so takes many Sockeye options. By default, Sockeye looks for a GPU. You can use a CPU instead by passing
paraphrase.sh --use-cpu
To change the beam or batch sizes, use the --beam-size X and --batch-size Y options.
You can set the GPU device with --device-ids N, where N is a 0-indexed CUDA device ID.
To get n-best output, use
paraphrase.sh --nbest-size K
Finally, you can experiment with negative and positive constraints by passing them in as the second and third tab-delimited field, respectively. Separate constraints are separated by a bar. For example,
echo -e "In times like this, one takes one’s happiness where one can find it.\tfortune|happiness|chances" | ./paraphrase.shwill use fortune, happiness, and chances as negative constraints. Positive constraints could be added with the use of a third field.