Rasa NLU to query App Store Top Charts

Related blog post with more in depth write up of the process used in this repo.

Output

The above image is a screen cap from a limited Plotly Dash app created to interact with the bot. The app code is in dash_demo_app.py.

The file live_test_rasa.py, allows a similar experience from the command line instead of via Dash in browser.

Input Data:

• data/generic_rasa_train_data.json: taken from the rasa intro restaurant chatbot example; all of the restaurant intent examples were removed
• data/app_chart_data.csv: table of top chart apps; the relevant column is just the list of app names to use as entities in training (table was created by utils/downloader.py)

Training process:

• Generated domain specific training data with gen_training_data.py and generic_rasa_train_data.json

  • Parameterized phrases were created to fill in the blanks with randomly chosen entities
    • eg: 'show me the {ordrank} most popular {chart} app'
  • Created N variations of the parameterized phrases and added them to the generic training data
  • output saved to data/app_train_data.json

• Train the rasa model

  • train_rasa.py (generic train script from rasa docs)