adtm/language-messenger-bot

Language learning Messenger ChatBot 👨‍🏫

A proof of concept of a Messenger Chatbot teaching new languages to users. The user is specified to choose a language and a scenario he will have to walk through. The chatbot can learn the user's behavior with AI.

Local development

• Download ngrok: https://ngrok.com/ or brew

• Set up environment variables:

  • Check .env.example and create .env with all the specified values in place
  • Project ID and credentials can be obtained from the dialogflow and Google Cloud Service dashboard. More on that: https://dialogflow.com/docs/reference/v2-auth-setup

• Run npm start / npm run dev and on an other terminal: ngrok http 1337 (port)

• Now paste this link to the Messenger Development platform

To add a "Get Started" button

curl -X POST -H "Content-Type: application/json" -d '{
  "setting_type": "call_to_actions",
  "thread_state": "new_thread",
  "call_to_actions": [
    {
      "payload": "Start"
    }
  ]
}' "https://graph.facebook.com/v2.6/me/thread_settings?access_token=<PAGE ACCESS TOKEN>"

language-messenger-bot's Issues

Plan next steps

1. Figure out what tools can we use from DialogFlow
2. Decide on what main functionalities we can build on these tools. Suggestions:
   1. Chatting with bot casually and it would correct your mistakes every time you typed or miss typed something badly instead of saying I don't understand. It could send a spelling card to you or the whole sentence written correctly. After that the conversation could continue.
   2. We should focus on maximum two languages first. It would be wise that bot could speak and understand both, simultaneously. Text translate pre built bot looks like it has some nice intents to build on that. In addition, we shallonly choose btw these four, as they are supported by small talk prebuilt bot (english, russian, french and italian)
   3. There are many pre-built bots which serve different purposes (like booking a restaurant). These real life situations (intents) could be borrowed, to help user practice the language.
   4. We could add bot responses on the go. That is, we would have all of them in one language, and constantly ask for the user to tell us how to say this phrase in our language. Then, when he would say we would add it to our response list.
3. Share work and do things. Still need to define a clear goal, so that we won't work in different directions.

Work on: Restaurant Booking bot

Need to fully implement restaurant booking bot. Prebuilt version mostly deals with "any" restaurant. Because we only want to simulate single specific restaurant's booking we need to modify the bot to resemble a real conversation.

We might use "required" inputs in order to ask user follow up questions until the restaurant is booked.

All of this might involve redefining the restaurant booking bot.

Implement routing to different agents

Link to gist which does something similar: "

https://bots.dialogflow.com/facebook/70ac5471-f9b0-4104-8f73-918eecc4f8a1/webhook

Link to an old forum thread, talking about he possible solution:

https://discuss.api.ai/t/how-to-switch-between-agents/6690/6

My suggestion is to use "main" bot as a proxy first, but then it would connect via hook to a backend service in case scenario is being played and then route the user to correct agent.

Proxy is needed because we still need to parse user input anyways.

There is also a thing that when the scenario is played out you don't want agent to recognize most of (or all) of main agent's (and, of course, all of the other agents') commands.

Test, how do agents handle false user input (misspeled words, etc)

The agent should at lest say that he does not understand what did user mean.

Ideally, it would even point to what exactly he did not understand.

Make the restaurant bot switch to main bot when the conversation is over.

Part of issue #15

Investigate other examples

Ivestigate other bot examples. Link: https://developers.facebook.com/docs/messenger-platform/getting-started/sample-apps

Bot should be able to deliver most used words in many languages, when asked for it.

Deliver ~10 most useful phrases in 10+ languages.

Set up heroku

Investigate Test Apps

Investigate NLP in Messenger Platform

Investigate how NLP is handled in Messenger Platform. Link: https://developers.facebook.com/docs/messenger-platform/built-in-nlp

Implement sender action for better user experience

Link: https://developers.facebook.com/docs/messenger-platform/send-messages/sender-actions.

Priority: Low

Set Up Travis or sth

Implement "main" agent, who would allow to choose which scenario user wants to try.

Subtasks:

1. Create an intro question, something like: "Hey! Which scenario would you like to try?" and then user would be presented with a list from possible scenarios. (or list alternative).
2. When user chooses a scenario, he should be routed to correct agent starting new conversation.
3. When conversation with a specific agent ends, user should be routed back.

Make a list (or sth) of possible scenarios in "main" bot

Right now, user is prompted for the scenario but he actually does not know what are the options. A list as a default response when scenario context is on, would solve a problem.

Make the "main" bot switch to restaurant bot

Part of issue #15

Implement a simple quiz, with multiple choice answers.

Chatbot should be able to deliver a quiz of ~5 different questions. Each question should be followed by an answer. After user answers, correct answer should be shown. Only one answer per quiz now.

May use this one: https://developers.facebook.com/docs/messenger-platform/send-messages/template/button