CMDPAD: Chinese Multimodal Dynamic Personality and Affect Dataset

🌟 Highlights

• From Recognition to Prediction: Moves beyond identifying current emotions to predicting the emotional trajectory of the next interaction turn.
• Dynamic Personality: Challenges the "static personality" assumption by providing dynamic utterance-level scores based on the Big Five traits using continuous sliding scales.
• High-Quality Multimodal Features: Includes aligned deep features for Text (T), Audio (A), and Vision (V).
• Experimental Validation: Results show that integrating personality traits with historical affect ("knowing who you are" + "knowing your recent mood") improves next-state prediction accuracy by 5%.

📊 Statistics & Source

• Source: 180 two-person dialogue clips from 60 popular Chinese TV dramas.
• Size: A total of 3,906 utterances.
• Modalities: Pre-extracted feature vectors are provided:
  • Text: bert-base-chinese
  • Audio: wav2vec2-base, wav2vec2-large-robust-emotion, wav2vec2-large-xlsr-chinese
  • Vision: vit-base, convnext-base (Vision was found to be the most critical modality for prediction tasks)

Data Split

Format	Train	Validation	Test	Total
Single-turn	2,734	585	587	3,906
Paired-turn	2,608	554	564	3,726

Current Hugging Face Release: The currently uploaded files contain only the training and validation splits.

📁 Dataset Structure

The dataset is split into two formats for single-turn and paired-turn (conversational context) tasks.

1. Single-turn Data

Used for training basic Affect Recognition (AR) and Personality Recognition (PR) models.

• multimodal_features_single.pkl: Contains name, mode (train/valid/test), and feature arrays for each modality.
• label_single.csv: Utterance-level annotations mapped by name.
  • Affect Labels: speaker_valence, valence_prediction (next-step prediction)
  • Dynamic Personality Labels: speaker_openness, speaker_neuroticism, speaker_extraversion, speaker_agreeableness, speaker_conscientiousness

2. Paired-turn Data

Used for training cascaded Affect Prediction in Conversations (APC) models.

• multimodal_features_paired.pkl: Contains paired features for Speaker A and Listener B (prefixed with A_ and B_).
• label_paired.csv: Paired labels for prediction tasks.

📥 Dataset Download

The data is protected and restricted to academic research. To request the dataset, please use an institutional (educational or research) email to send the signed User License Agreement (ULA) to [hengxie@bit.edu.cn], then apply on Hugging Face; the download will be authorized upon verification.

💻 Baseline Code

The official PyTorch implementation for this work is available here:

• GitHub Repository: https://github.com/HensonXie/CMDPAD

📌 Release Note

The test split is reserved for a forthcoming academic competition and will be made publicly available after the competition has concluded.

🤝 Citation

If you find this work useful for your research, please cite our paper (Accepted by Pattern Recognition):

@article{zhou2026cmdpad,
  title={CMDPAD: A Chinese multimodal dynamic personality and affect dataset for affect prediction in conversations},
  author={Zhou, Zisen and Xie, Heng and Wen, Chang and Liu, Xuefei and Tao, Jianhua and Wen, Zhengqi and Li, Changsheng and Lian, Zheng and Zhao, Jinming and Xiong, Bingsen and Qin, Shaozheng},
  journal={Pattern Recognition},
  pages={113822},
  year={2026},
  publisher={Elsevier}
}