Code for "Few-shot Dialogue Strategy Learning for Motivational Interviewing via Inductive Reasoning"

• File diretory structure:

  • train_agent.py
    • for learning dialogue strategies
  • inference.py
    • for evaluating dialogue strategies
  • tom_detector.py
    • util functions for indexing experiences (strategies) with user mental state, that inferred user mental state
  • utils
    • building blocks, should be readable
  • AnnoMI
    • The dataset (also, the code we used to split the data)

• Getting Started

  • Fill out the assets.py in appendix, it needs
    • your OPENAI api key
    • path to a huggingface-compatible classifier (like BERT, see below)

• Obtaining the dialogue act classifier:

  • Use Huggingface Trainer to finetune mental-bert on dataset by Welivita et. al. We report hyper-parameters in the appendix (and feel free to use your favorite fine-tuning codebase).
  • Unless otherwise noted in our paper, all parameters are default parameters.