TPol - Translate First Reorder Later

This is the repo for paper EACL23 Findings paper.

Installation

First, create a Conda environment using the tpol-environment.yml file and activate it using the following commands:

conda env create -f tpol-environment.yml
conda activate tpol-private

Instructions

1. Import the data: clone the GEO-Aligned repository in data/:

$ cd data/
$ git clone https://github.com/interact-erc/GEO-Aligned.git
$ cd ..

2.Run the Lexical Translator script. You have the option to choose between the bert and the mbart approach. For instance, to run the bert approach, execute the bert_translator.pyz script. 3. Run the MR Reorderer script. Similarly, you can choose between the bert and the mbart approach. For instance, to run the mbart approach, execute the mbart_reorderer.py script.

Both the translator and reorderer scripts require the following arguments:

- --dataset: path to the dataset file
- --test-ids: path to the test ids file
- --val-ids: path to the dev ids file
- --language: dataset language, choose among en, it, de
- --out-file: path to the file where the test predictions are be saved
- --results-file: path to the file where the numerical results are saved

Additionally, the reorderer script requires the --lexical-predictions argument, which should be the file outputted by the translator --out-file.

To run mbart_reorderer in silver mode, first run the translator with the additional --all-predictions-file argument (the path to the file storing the predictions). Then, provide the same file to the reorderer as the --silver-predictions-path.