Talk to Joey: How to deploy a Joey NMT model as a slack translation bot

This is a quick guide on how to locally deploy a trained Joey NMT model as a slack bot. It's a great way to get a good feeling for what it has learned and a very simple way to show off your model without implementing a front-end.

Disclaimer: Not made for long-term or production-ready deployment, since we're not using a "proper" webservice.

Requirements

• You need a trained Joey NMT model. See here for instructions on how to train one, or use one of the pretrained models.
• In order to provide translations, Joey NMT needs to be running on a machine. GPUs are faster, but the reponse time of a model running on CPU should still be bearable (imagine a human typing - it's still faster than that ;)). You can only query translations as long as the job on the machine is running, so best would be a server. For short-time demos, your local machine should be fine, too.
• This code runs on Python3.6.
• We're using ngrok to expose a locally deployed model to the outside.

Setup

Installation

Install the required packages:

python3.6 -m pip install -r requirements.txt

This includes Joey NMT (for CPU - install it manually for GPU, see next section).

Joey NMT

• You need to install Joey NMT and its requirements first, see here.
• Train a model. Let's assume it's stored in my_model_dir. This directory should contain at least one checkpoint, the vocabularies and the configuration file.

1. Create an app

• Create or choose a channel in your slack team to integrate your bot. This is where the bot reacts to every incoming message by anyone. Let's call this channel BOT_CHANNEL. Write it's name in bot.channel.
• Create an app for your workspace. See this tutorial.
  • Write the app's name (BOT_NAME) into bot.name.
  • Define bot token scopes: app_mentions:read, chat:write, incoming-webhook, channels:read are needed. You need to re-install the app anytime you change the permissions.
  • Authorize the app for the workspace and assign it the new channel.
  • If this is successful, you'll receive a bot token. It should start with xoxb.
  • Write the bok token into bot.token.
  • Add the bot to the channel in the slack workspace.
  • Store the sign-in secret in bot.signin.
• We're going to use the Event API to make the bot subscribe to events in slack. (The RTM API is no longer available for new apps.)

2. Subscribe to events

• Install ngrok. It will allow us to expose a local service to the public.
• Start ngrok on a port 3000: ./ngrok http 3000. In order to interact with your app, this process needs to be running.
• Copy the url that ngrok reports. It should look like http://somerandomsymbols.ngrok.io.
• Enable Event Subscriptions for your app, as described here.
• For the Request URL, use the ngrok URL with a suffix: http://somerandomsymbols.ngrok.io/slack/events.
• Run python3.6 main.py my_model_dir to start the app.
• After verification, subscribe to bot events: app_mention, message.channels, message.im.
• Click Save changes to save the changes for your app.

3. Running

• Run python3.6 main.py my_model_dir to start the app. Optionally specify --bpe_src_code my_model_dir/src.bpe if you're using a BPE-based model, and --tokenize if the training data was tokenized too.
• In slack, move to the BOT_CHANNEL and write a message. Your bot should automatically reply.
• In addition to that, the bot reacts on mentions, so addressing @BOT_NAME will make it respond.

Configurations and Customization

• Make sure to edit the config.yaml in the model_dir according to your use case. Mind the setting for the following:
  • use_cuda: set to False if running on CPU, True when on GPU.
  • max_output_length: sets the maximum output length
  • beam_size: beam size for decoding, 1 is greedy decoding.
• We assume the data is pre-processed with the MosesTokenizer if you set --tokenize. If you want to use a different one, modify the code accordingly.

Interactions

See the docu on how to create bots for slack workspaces and explore the links there: https://slack.com/help/articles/115005265703-Create-a-bot-for-your-workspace You could add more interaction modes, language id to activate different bots, etc. - please make a PR to this repo if you implement a cool extension :) This is really just the bare bone.