CS 337 - Remy: A Conversational Recipe Chatbot

This project aims to provide a conversational interface for interacting with the recipe transformer. We utilize the Rasa contextual AI assistant framework to handle user input, determine their intent, and carry out the proper actions when interacting with our recipe transformer. Additionally, we were able to integrate this code with Slack in order for the user to be able to interact with this agent in a Slack workspace.

Team Members (Group 4): Cameron Barrie, Alexander Reneau, Marko Sterbentz

The repository for this project can be found on GitHub: https://github.com/MarkoSterbentz/nlp-recipe-chatbot

Required Packages

Rasa
- Installation pip install rasa
Pandas (for data containers)
- Installation: with pip install pandas
Spacy (for general NLP tasks)
- Installation: pip install -U spacy
- We use the en_core_web_sm model, which can be installed with python -m spacy download en_core_web_sm after installing Spacy.
BeautifulSoup (for web scraping)
- Installation: pip install beautifulsoup4

Note: A full list of all packages installed can be found in requirements.txt.

Installation

Note: The following installation instructions were tested on MacOS 10.14, and assumes that Anaconda is installed. This is primarily needed if you are having difficulties installing Rasa or any of the packages within an existing environment.

Setting up Rasa

It is recommended that you start by setting up a new virtual environment. We do this through Anaconda with the following command:

conda create -n <env_name>

Then, activate this environment so that we can start installing packages there:

source activate <env_name>

Install pip in this environment:

conda install pip

Switch to a Python version in this environment that is compatible with Tensorflow, and thus Rasa as well:

conda install python=3.7.5

Install rasa:

pip install rasa

Alternatively, you can also run pip install -r requirements (instead of installing Rasa on its own) in the base project directory to install everything that is required for this project, including Rasa.

Running the Terminal Version of the Project

This project assumes the use of Python 3.7. From the base project directory, make sure you have the proper virtual environment activated and run the following command to start the Rasa action server:

rasa run actions

Then, in another terminal window, run the following command to ensure there is a model for Rasa to use:

rasa train

Once it is done training a model, run the following command in the same terminal to start interacting with Remy:

rasa shell

Connecting the Project to Slack

Note: We primarily followed this tutorial when connecting our agent to a Slack workspace. Please follow the instructions there to properly configure your workspace and allow for interaction with Remy.

Once you've finished configuring the Slack workspace following the instructions at the link above and have ngrok installed, run the following command:

rasa run actions

Then, run ngrok in a different terminal to create the tunnel from the local server to Slack:

ngrok http 5002

Finally, in yet another terminal window, run the following command to start Remy:

rasa run --port 5002 --connector slack --credentials slack_credentials.yml --endpoints endpoints.yml

Lastly, go back to the app settings for Remy (go to https://api.slack.com/apps/ and click on Remy), click on Event Subscriptions on the left sidebar, and input the updated ngrok url (it should look something like http://97f49c33.ngrok.io/webhooks/slack/webhook).

With all of this done, reload the Slack workspace. Then, you should be able to click on Remy in the left sidebar, click on the Messages tab,and start talking to Remy about recipes. An example of this interaction is shown in the screenshot below:

How it Works

This project builds off of the previous course project in order to provide a nice conversational interface for working one's way through a recipe. We utilized Rasa for parsing in input from the user, as well as extracting intent from their statements. These intents are then mapped to user interaction stories, which can be found in stories.md. Rasa's NLU module evaluates the user's input to predict where the interaction is going, and uses this prediction to provide the appropriate prompts/output and to carry out the required actions for the user.

The full domain for this conversational agent can be found in domain.yml.

Once Remy is running, it can be instructed to perform a variety of actions:

Retrieve and parse a recipe from AllRecipes.com
Display the ingredients of the recipe
Display all of the cookings steps in the recipe
Display the current cooking step in the recipe
Go to a specific cooking step in the recipe
Provide links to video answers for 'how to' questions
Provide links to web searches for 'what is' questions

Current Shortcomings

Currently, the parsing and language understanding present in our model is not very good. This is due to insufficient NLU training examples. We plan to expand the training examples within nlu.md in order for Remy to be able to handle a wider variety of user input.