GEM-Benchmark 2021 Shared Task Submission

Code release for submission made to GEM-Benchmark 2021 Text Simplification Shared-Task on TurkCorpus ans ASSET datasets

Getting Started

Dependecies

• Python >= 3.7

Installation

git clone https://github.com/kvadityasrivatsa/gem_2021_simplification_task.git
cd gem_2021_simplification_task
./install.sh

How to use

Train the submission model on WikiLarge

• for TurkCorpus:

python3 train.py --evalset turk --ner --nbchars 0.95 --levsim 0.75 --wrdrank 0.75

• for ASSET:

python3 train.py --evalset asset --ner --nbchars 0.95 --levsim 0.75 --wrdrank 0.75

Generate and evaluate output (on SARI score)

• for TurkCorpus:

python3 evaluate.py --evalset turk

• for ASSET:

python3 evaluate.py --evalset asset

Pretrained Model

The checkpoint for our model with the best scores is available here

Model Desciption

(Note: The official system-desciption for the model can be found here)

Our model builds upon the ACCESS model proposed in Controllable Sentence Simplification (Martin et al., 2020).

Authors

• KV Aditya Srivatsa (k.v.aditya@research.iiit.ac.in)
• Monil Gokani (monil.gokani@research.iiit.ac.in)

If you have any queries, please do reach out.

License

Refer to the LICENSE file for more details.