Integrate machine translation to Darma chat

[thammegowda]

Patch description

• Enables two-way machine translation for MT
• Configurable MT backends: RTG, HF transformers with flexible argument support
• Flexibility of translating human input, bot output, both, or None

Testing steps

cd parlai/crowdsourcing/tasks/darma_chat 
python run.py

Other information
requires installation of transformers lib

---

( docs copied from README)

Enabling MT

see translator.py for the code

Add this config block as mephisto.blueprint.translator

translator:
  activation: 'pre' # pre, post, pre+post, null
  preprocess: rtg_api
  preprocess_args:
    # TODO: change the URL to DARMA hosted service
    api_url: http://rtg.isi.edu/many-eng/v1/translate
  postprocess: huggingface
  postprocess_args:
    model: Helsinki-NLP/opus-mt-en-fr

The key activation takes the following values

• pre - Only translate human input (via preprocess config)
• post - Only translate bot output (via postprocess config)
• pre+post - Enable both pre and post
• null - Disable MT. Which has same effect as deleting the whole translator config block

preprocess and postprocess takes the MT backend name.
Whereas {pre,post}process_args take a dictionary of arguments to MT backend.

The following MT backends are supported

• rtg_api which calls RTG over a REST API. See http://rtg.isi.edu/many-eng/
• huggingface calls transformers library. Requires model argument which can be obtained from https://huggingface.co/models?pipeline_tag=translation