Brain in a Jar, or Modern Virtual Assistants: Exploring the Evolution of NLU in Virtual Assistants

This is a repository with code for presentation about Virtual Assistants I gave on PyData Bydgoszcz in November 2023.

Link to presentation

Joint NLU Live Coding

Follow these steps to reproduce the Natural Language Understanding (NLU) model training from our recent PyData event session.

1. Environment Setup

   • Create a Conda environment:

     conda create --name joint_nlu python=3.10
     

   • Activate the environment:

     conda activate joint_nlu
     

   • Install dependencies:

     pip install -r requirements.txt
     

     Alternatively, this step can be integrated into the environment creation process.

2. Grammar Modification

   • Navigate to the grammars directory:

     cd grammars/
     

   • Modify the calendar.gram file as desired.

3. Grammar Expansion

   • Expand the grammars into corpora:

     python generate_patterns.py -f calendar.gram > iva-calendar-0.1.0.tsv
     

   • Fix two common errors:
     1. Remove double underscores ("__"). This can be done manually or using sed:

        sed -i 's/__//g' calendar.gram
        

     2. Eliminate spaces surrounding tabs. Replace (space)(tab)(space) with (tab).

4. Slot Expansion

   • Run the slot expansion script:

     ./expand-slots.sh iva-calendar-0.1.0.tsv iva-corpus-calendar-0.1.0.tsv true true
     

5. Configuration Adjustments

   • Modify the join_nlu configuration in joint_nlu/calendar.json. Experiment with the num_train_epochs setting as needed.

6. Corpus File Naming

   • Ensure your corpus file name matches the _URL in custom.py.

7. Model Training on Google Colab

   • Import joint_nlu_train_on_colab.ipynb into Google Colab and follow the instructions therein.
   • After cloning the joint_nlu repository in Colab, upload three files into the joint_nlu directory: calendar.json, the corpus file, and custom.py.
   • Proceed with the training.

Upon successful completion, assuming "push_to_hub": true is set in calendar.json, your model will be uploaded to your Hugging Face Hub. For deployment and usage, refer to the Joint NLU Repository.