CSO: Chain of Strategy Optimization Makes Large Language Models Better Emotional Supporter

This repository contains code for the EMNLP 2025 Findings paper Chain of Strategy Optimization Makes Large Language Models Better Emotional Supporter

ESCPro Dataset link

Overview

The system generates high-quality emotional support conversations through a three-stage pipeline:

1. Conversation Tree Generation (run.py): Uses MCTS to explore different emotional support strategies and generate conversation trees
2. Preference Data Construction (build_data.py): Converts conversation trees into preference-annotated dialogue histories
3. Dataset Conversion: Transforms dialogue histories into either preference pairs (change_data.py) or KTO datasets (change_data_kto.py)

Pipeline

Stage 1: Generate Conversation Trees

python run.py

• Uses MCTS to explore emotional support strategies
• Generates conversation trees with multiple response options
• Saves trees as pickle files in output/tree/

Stage 2: Build Preference Data

python build_data.py

• Converts conversation trees into dialogue histories
• Extracts preferred vs. non-preferred responses
• Generates comparison data with strategy annotations

Stage 3: Convert to Training Datasets

For Preference Pairs:

python change_data.py

For KTO Dataset:

python change_data_kto.py

Citation

@article{zhao2025chain,
  title={Chain of Strategy Optimization Makes Large Language Models Better Emotional Supporter},
  author={Zhao, Weixiang and Sui, Xingyu and Han, Xinyang and Deng, Yang and Hu, Yulin and Guo, Jiahe and Qin, Libo and Du, Qianyun and Wang, Shijin and Zhao, Yanyan and others},
  journal={arXiv preprint arXiv:2503.05362},
  year={2025}
}