This repository contains code for our paper DEMO: Reframing Dialogue Interaction with Fine-grained Element Modeling. We systematically construct the dialogue framework from the Prelude through the Interlocution to the Epilogue and define an innovative research task: Dialogue Element MOdeling. Furthermore, we introduce a tailor-designed benchmark DEMO to facilitate comprehensive dialogue modeling and assessment. Concretely, our proposed task focuses on two core competencies of models: (1) Element Awareness, which entails reverse engineering to decompose dialogue elements, and (2) Dialogue Agent Interaction, which involves goal-directed multi-turn dialogue modeling driven by elements. We meticulously design a data synthesis framework, contributing to a novel benchmark for dialogue modeling that encompasses multifaceted elements applicable to both English and Chinese. Besides, inspired by imitation learning, we amass a substantial collection of expert experiences and build a DEMO agent endowed with dialogue element modeling.

• [2024.12.07]🔥DEMO is coming! We release the paper, code, models, and data for dialogue element modeling!

Step1 Download DEMO and unzip data

git clone https://github.com/MozerWang/DEMO.git
cd DEMO
unzip data/DEMO.zip -d data/

Step2 Create a conda environment and Install other dependencies.

conda create --name loong python=3.9 -y
conda activate DEMO
pip install -r requirements.txt

Step3 Preparing the Model

1. (Must) Set up your OPENAI key in config/gpt_4o.yaml

api_key: "Your OPENAI key"

2. If you are using API-based LLM

# Firstly, Set up your key in config/*.yaml
api_key: "Your API key"

3. If you are using Open-sourced LLM

# We recommend using vLLM. And we use HTTP server that implements OpenAI’s Completions and Chat API.
# Set up your vLLM settings in config/*.yaml

Step4 Evaluate

sh run.sh

@article{wang2024demo,
      title={DEMO: Reframing Dialogue Interaction with Fine-grained Element Modeling}, 
      author={Minzheng Wang and Xinghua Zhang and Kun Chen and Nan Xu and Haiyang Yu and Fei Huang and Wenji Mao and Yongbin Li},
      year={2024},
      journal={arXiv preprint arXiv:2412.04905},
      url={https://arxiv.org/abs/2412.04905}
}