OpenTIPE: An Open-source Translation Framework for Interactive Post-Editing Research

OpenTIPE is a flexible and extensible framework designed to support research on interactive post-editing. With an interactive environment, OpenTIPE allows researchers to explore human-centered approaches to the post-editing task.

In order to see the tool in action, we encourage the reader to take a look at the following resources:

• Two minute walkthrough video
• Interactive online demo

Deployment using docker-compose

• Start test server:

  docker compose up --build

• Start the automated tests:

  docker compose -f docker-compose.yml -f docker-compose-tests.yml up --build --abort-on-container-exit

Configuration

Backend/General

Any configuration can be done in the .env file prior to starting docker-compose.

The following configurations are supported:

Env variable	description
MOCK_MODEL	If True, the models will return a predefined string instead of running a model
HARD_CODED_RESPONSE	If True, the models will return a hard coded string, if False it will return the input text
APE_MODEL_URL_DIR	The URL to the productive APE model directory
MT_MODEL	Choose between DeepL, HuggingFace & FairSeq implementations of the machine translation model
MT_MODEL_NAME	If HuggingFace or FairSeq are chosen for the MT_MODEL, choose an applicable model checkpoint. HF Examples: Helsinki-NLP/opus-mt-de-en | (Helsinki-NLP/opus-mt-ine-ine) Fairseq Options: transformer.wmt19.de-en | transformer.wmt14.de-en
DEEPL_API_KEY	This Variable must be exposed in your environment files
AUTH_ENABLED	Enable/disable Firebase bearer token authentication by setting this variable to True/False respectively
GPU	Enable/disable GPU support by setting this variable to 0/-1 respectively

Frontend

Any frontend-related configuration can be done in frontend.env. Please refer directly to the file to see the available variables and their descriptions.

CPU Only deployment

In case you want to run OpenTIPE using docker with a GPU, perform the following actions:

• Set GPU=0 in the .env file as explained under Configuration
• Uncomment the following block from the docker-compose.yaml file:

  deploy:
    resources:
      reservations:
        devices:
          - driver: nvidia
            count: 1
            capabilities: [gpu]

  This block is used to give a service access to the GPU but raises an error if there is no GPU available.

Using custom models

There are two ways to make this project run with your own models:

Use existing architecture and add custom weights

For licencing reasons we cannot provide our trained APE models. However, you can provide your own model files to the APE service in order to use the provided APE model architecture as follows:

• In the .env file, set MOCK_MODEL=False
• Create a directory in ./model_data for each model that you want to make available, e.g. ape-model-huggingface-de-en.
• Inside the newly created directory, place the following files:
  • config.json
  • pytorch_model.bin
  • run_config.json
  • special_tokens_map.json
  • tokenizer_config.json
  • vocab.txt

When you next start the application using docker compose, the machine translation (MT) and automatic post editing (APE) models should be loaded.

If you don't have access to these files, read the README.md in the ape-model directory in order to train the model.

Add custom model code

If you want to add your own model code and weights, create your own service, analog to either mt-model or ape-model or simply edit the existing code for these services.