Complexity Controlled MT

This repository contains implementation for

Sweta Agrawal and Marine Carpuat. "Controlling Text Complexity in Neural Machine Translation". EMNLP-IJCNLP 2019.

Usage Instructions

1. Run bash setup.sh to obtain necessary software.
2. You can request the Newsela dataset and the OPUS corpus from here and here respectively.
3. Use PrepareDataFiles.ipynb to extract alignments and create dev/train/test split and put it under data/.
4. For Training

Monolingual English Simplification 
> bash scripts/main.sh -d -i <iter>

Multitask model trained on English Simplification and out-of-domain bilingual data
> bash scripts/main.sh -d -b -i <iter>

Other options:
> bash scripts/main.sh -h

5. Evaluation: You can evaluate on the complexity controlled MT task by using the c flag

> bash scripts/evaluate.sh -i <iter> -c

Other options:
> bash scripts/evaluate.sh -h

To evaluate the pipeline model, you need to supply the ids for the first and the second model and a name for the evaluation. So if 1 represents the translation model (ES-EN) and 2 represents the simplification model (EN-EN), you can use the following script to evaluate the pipeline translate-then-simplify.

> bash scripts/evaluate_pipeline.sh -i 1 -j 3 -k esen

Some of the scripts for this work were modified from here. Also, the scripts to calculate SARI score and ARI readability index were referred from here and here respectively.

Note: Please refer to the supplementary material of our paper (Table 10 and 11) for exact statistics of the dataset used for training.