Dialogue Systems for Emotional Support via Value Reinforcement

This is the official repository for the paper:
"Dialogue Systems for Emotional Support via Value Reinforcement"

---

🚀 Training

The framework consists of three key components:

1. Target Value Detector
   Identifies which human values to reinforce at each turn.

2. Reference Generator
   Generates utterances that promote these values from the seeker.

3. Supporter Model
   Determines appropriate strategies and responses based on values and references.

✅ Target Value Detector

• Training Command

  bash tvd_sft.sh

✅ Reference Generator

• Training Command

  bash rg_sft.sh
  bash rg_dpo.sh

✅ Supporter Model

• Training Command

  bash sptr_sft.sh
  bash sptr_dpo.sh

---

🧪 Simulation

After training all three components (target value detector, reference generator, and supporter model), you can simulate a conversation with the seeker simulator using the seeker personas (test dataset).

• Run Simulation

  bash simulation.sh