artursstaf/dynamic-learning

Marian NMT dynamic adaptation example

This repository aims to provide working self-contained example of Marian NMT system with dynamic adaptation.

The example builds and runs as 3 Docker containers

• Marian adaptive server
• Translation memory service
• Wrapper service

Marian NMT dynamic adaptation works by fine-tuning model on context - (source, target) sentence pairs. So for each translation Marian requires the input text and context for fine-tuning.

Translation memory provides the context sentences for Marian. The current implementation: https://github.com/tilde-nlp/lucene-sentence-search/ Provides with functionality to:

• Store indexed (source, target) sentence pairs.
• Retrieve similar sentences by BLEU.
• Drop sentence pairs.

Wrapper is just an additional service that simplifies API by forwarding calls to Translation Memory and Marian. Wrapper service also serves as a simple example how Translation Memory and Marian interacts.

Prerequisites

The example is built and run with Docker and docker-compose v1.28.0+ and is expected to run on a machine equipped Nvidia GPU https://docs.docker.com/compose/install/

How to build and run example

First download model files by running bash download_model.sh in the repo root directory.
Then build and start the services by docker-compose up.
Which will then load the downloaded marian model and expose the API on port 8088.

Configuration

Wrapper, Marian and Translation Memory startup parameters are defined in docker-compose.yml By default it's configured to load downloaded example marian model in ./marian-model and run on GPU: 0.

API

Wrapper service exposes 3 actions:

• /translate Translate input sentence, context is fetched from TM.
• /save Save (source, target) sentence pair in TM by user ID.
• /delete Drop all saved sentences in index by user ID.

Wrapper API call examples:

POST http://localhost:8088/translate
Content-Type: application/json

Request: 
{
  "input": "Hello world !",
  "meta": {
    "uid": "Artūrs",
    "srclang": "en"
  }
}

Reply:
{
  "output": "Sveikas pasaulis !"
}

POST http://localhost:8088/save
Content-Type: application/json

Request:
{
  "source": "Hello world !",
  "target": "Sveika pasaule !",
  "meta": {
    "uid": "Artūrs",
    "srclang": "en"
  }
}

Reply: 
{
  "errorMessage": null,
  "status": "OK"
}

POST http://localhost:8088/delete
Content-Type: application/json

Request:
{
    "uid": "Artūrs"
}

Reply: 
{
  "errorMessage": null,
  "status": "OK"
}

Using different Translation Memory

Marian itself does not depend on a specific Translation Memory implementation. Different ways of acquiring context can be employed. An example implementation of how marian-adpative server is called can be found in wrapper_service/wrapper_service.py

ws://localhost:80/translate

Sent:
{
    "input": "Hello world !",
    "context": [
        "Source context sentence 1\nSource context sentence 2\nSource context sentence 3",
        "Target context sentence 1\nTarget context sentence 2\nTarget context sentence 3"
    ]
}

Received:
{
  "output": "Sveikas pasaulis !"
}