This repository provides an implementation of the joint RAG-based framework introduced in the paper "Interpersonal Memory Matters: A New Task for Proactive Dialogue Utilizing Conversational History" . The framework is designed to tackle the Memory-aware Proactive Dialogue (MapDia) Task , which focuses on utilizing conversational history to generate proactive and contextually relevant dialogues.

Interpersonal Memory Matters: A New Task for Proactive Dialogue Utilizing Conversational History

The project includes code for the inference process of the proposed framework, demonstrating how to combine a search module, ranker, and proactive dialogue model to generate proactive responses based on conversational history.

The required test data can be downloaded from the Huggingface Dataset repository:
FrontierLab/ChMap-Data.

While the repository does not include training scripts, you can driectly download data from our Huggingface repository to conduct data for training Topic Summary Model, Topic Retrieval Model, and Memory-Aware Proactive Response Generation Model. The required test data can be downloaded from the Huggingface Dataset repository: FrontierLab/ChMap-Data .

• Implements the inference pipeline for the MapDia task.
• Supports preprocessing of test data to extract key fields, reducing runtime overhead when running the full pipeline.

The test data required for inference can be downloaded from the Huggingface Dataset repository: FrontierLab/ChMap-Data.

Download the overall_dialogue_review/test.json file and place it in the root directory of this project.

To reduce runtime overhead during inference, preprocess the test data using the test_data.py script:

python test_data.py

This script extracts the necessary fields from the raw test data and saves them in a simplified format for faster processing.

Once the test data is preprocessed, you can run the inference pipeline such as:

python ours.py

The script will generate proactive dialogue responses based on the provided conversational history.

If you find this project useful, please cite the original paper:

@misc{wu2025interpersonalmemorymattersnew,
      title={Interpersonal Memory Matters: A New Task for Proactive Dialogue Utilizing Conversational History}, 
      author={Bowen Wu and Wenqing Wang and Haoran Li and Ying Li and Jingsong Yu and Baoxun Wang},
      year={2025},
      eprint={2503.05150},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.05150}, 
}

For any questions or issues, please open an issue in this repository. Contributions are welcome!