PMPC: Prompt Learning with Multiperspective Cues for Emotional Support Conversation Systems

Reproduction of the following paper:

Yangyang Xu, Zhuoer Zhao, Xiao Sun, and Xun Yang. "Prompt Learning With Multiperspective Cues for Emotional Support Conversation Systems." IEEE Transactions on Computational Social Systems, 2025. [paper]

@article{xu2025pmpc,
  title={Prompt Learning With Multiperspective Cues for Emotional Support Conversation Systems},
  author={Xu, Yangyang and Zhao, Zhuoer and Sun, Xiao and Yang, Xun},
  journal={IEEE Transactions on Computational Social Systems},
  year={2025},
  doi={10.1109/TCSS.2025.3539915}
}

Overview

PMPC extracts multiperspective cues from dialogue history and constructs soft prompts to guide a BlenderBot-small backbone for emotional support response generation.

Cues Catcher (Section III-A):

• Topic Cues (Ht): GloVe + PMI-based keyword extraction
• Prior Knowledge Cues (He): DPR retrieval from training responses
• Mental State Cues (Hc): COMET-ATOMIC 2020 (6 user + 3 listener relations)
• Other Cues: Post (Hp), Situation (Hs) from encoder hidden states

Prompt Builder (Section III-B):

• Semantic enhancement prompt (Pe) for encoder
• Semantic constraint prompt (Pd) for decoder, conditioned on strategy prediction

Base Code

This implementation builds upon the ESConv codebase:

Siyang Liu*, Chujie Zheng*, et al. "Towards Emotional Support Dialog Systems." ACL 2021. [paper] [repo]

Setup

Environment

conda create -n pmpc python=3.9
conda activate pmpc
pip install torch transformers nltk numpy tqdm

External Models

Download and place under external_models/:

Model	Source	Path
COMET-ATOMIC 2020	HuggingFace	external_models/comet-atomic-2020/
DPR Context Encoder	HuggingFace	external_models/dpr-ctx-encoder/
DPR Question Encoder	HuggingFace	external_models/dpr-question-encoder/
GloVe 6B 300d	Stanford NLP	external_models/glove.6B.300d.txt

Also download BlenderBot-small-90M into Blenderbot_small-90M/.

Data Preprocessing

cd _reformat && python process.py && cd ..
bash RUN/prepare_pmpc.sh

Training

bash RUN/train_pmpc.sh

Inference

bash RUN/infer_pmpc.sh

Select the best checkpoint from DATA/pmpc.pmpc/ based on validation PPL, and set --load_checkpoint accordingly in the inference script.