Introduction

This code repository for An Energy-based Model for Word-level AutoCompletion in Computer-aided Translation.

Project Structure

.
├── criterions  # Directory for custom Fairseq loss functions
├── data        # Directory for custom Fairseq data classes
├── fairseq_generate.py  # Modified file based on fairseq.generate
├── generate_ar.py       # Modified file based on fairseq.generate
├── generate.py          # Modified file based on fairseq.generate
├── __init__.py
├── interactive.py       # Modified file based on fairseq.interactive
├── models       # Directory for custom Fairseq model classes
├── modules      # Directory for custom modules
├── preprocess.py # Modified file based on fairseq.preprocess
├── README.md
├── requirements.txt
├── scripts      # Directory for various script files
├── tasks        # Directory for custom Fairseq tasks
├── train.py     # Modified file based on fairseq.train
├── validate.py  # Modified file based on fairseq.validate

Local Installation Environment

pip install -r requirements.txt

Fairseq version used

https://github.com/yc1999/fairseq.git

This version of Fairseq is based on the official version '1.0.0a0+c8d6fb1'. The only modification is in the fairseq/models/transformer/transformer_config.py file, with no other changes.

Other packages, such as torch, should be installed separately. You can refer to the requirements.txt file in this directory for installation.

Quick Start

1. Set the following three environment variables in the shell command line:

   export save_root=${YOUR_PROJECT_PATH}/save
   export project_root=${YOUR_PROJECT_PATH}
   export dataset_root=${YOUR_PROJECT_PATH}/dataset  
   

2. Train the model using the following command:

   # zh-en
   bash ${project_root}/scripts/rerank/zh-en/start_train.sh bi_context
   # bash ${project_root}/scripts/rerank/zh-en/start_train.sh prefix
   # bash ${project_root}/scripts/rerank/zh-en/start_train.sh suffix
   # bash ${project_root}/scripts/rerank/zh-en/start_train.sh zero_context
   
   # en-zh
   bash ${project_root}/scripts/rerank/en-zh/train_pretrained.sh bi_context
   
   # de-en
   bash ${project_root}/scripts/rerank/de-en/train_pretrained.sh bi_context
   
   # en-de
   bash ${project_root}/scripts/rerank/en-de/train_pretrained.sh bi_context
   

3. After training is completed, run the following command to infer the model:

   bash ${project_root}/scripts/rerank/zh-en/start_valid.sh
   

   Note: Please modify the ckpt directory in valid.sh as needed.