CCPD — Classical Chinese Poetry Theme & Emotion Classification

Multi-label joint theme and emotion classification of classical Chinese poetry for PoetryMTEB embedding evaluation.

Upstream CCPD provides 17,103 poems with multi-label theme and emotion annotations, introduced with a knowledge-guided Transformer model (Wei et al., IEEE/ACM TASLP 2024).

Dataset Card

Item	Description
Source	yuting-wei/CCPD
Paper	Wei et al., Knowledge-guided transformer for joint theme and emotion classification of Chinese classical poetry, IEEE/ACM TASLP 32 (2024) 4783–4794. DOI 10.1109/TASLP.2024.3487409
Languages	Classical Chinese / Chinese (zh)
Unit	Full poem text (poem)
Labels	Multi-label 10 themes + 12 emotions (joint annotation)
Size	train=10071; test=7032; total=17103
Splits	Preserved from the previous PoetryMTEB/CCPD release (≈59% / 41%)
License	CC BY 4.0 (PoetryMTEB redistribution; cite original authors)
Evaluation metrics	Multi-label classification on embeddings: macro/micro F1, Average Precision (AP) (themes and/or emotions). Upstream paper reports joint theme–emotion classification metrics.

Label taxonomy

Themes (10)

id	label_name	name_en	train	test	total
0	思乡	Homesickness	5476	3534	9010
1	咏物	Object chanting	335	231	566
2	山水	Landscape	603	468	1071
3	怀人	Longing for someone	1363	900	2263
4	悼亡	Elegy / mourning	306	270	576
5	怀古 (paper: 咏史)	Historical reflection	377	312	689
6	田园	Pastoral	710	503	1213
7	战争 (paper: 边塞战争)	Frontier / war	611	528	1139
8	闺怨	Boudoir lament	361	289	650
9	送别	Farewell	144	170	314

Emotions (12)

id	label_name	name_en	train	test	total
0	愁绪	Melancholy	1830	1718	3548
1	哀伤	Sadness	2034	1611	3645
2	恐惧	Fear	522	411	933
3	惊讶	Surprise	298	211	509
4	流泪	Tears	562	392	954
5	失意	Frustration	908	668	1576
6	思念	Yearning	869	599	1468
7	孤独	Loneliness	938	708	1646
8	喜悦	Joy	314	208	522
9	怨恨	Resentment	394	288	682
10	愤怒	Anger	135	87	222
11	想家	Homesick feeling	7396	4931	12327

Codebook: label_taxonomy.json.

Name aliases vs paper: this packaging keeps Hub names 战争 / 怀古 corresponding to paper 边塞战争 / 咏史. Rare typo 哭泣 (1 train example) is merged into 流泪.

Features

Field	Type	Description
id	string	Example id (ccpd-{split}-{idx})
title	string	Poem title
poem	string	Classification input: full poem body (formerly text)
theme_labels	list[int64]	Theme class indices
theme_label_names	list[string]	Theme names
emotion_labels	list[int64]	Emotion class indices
emotion_label_names	list[string]	Emotion names

Construction method

1. Load existing PoetryMTEB/CCPD train/test parquet.
2. Normalize fields: text→poem; map theme/emotion strings to stable taxonomies + integer ids.
3. Apply paper-name aliases and merge 哭泣→流泪.
4. Write PoetryMTEB parquet under data/, plus dataset_infos.yaml and Dataset Card.

Split membership is unchanged from the prior Hub release (total still 17,103).

How to load

from datasets import load_dataset

ds = load_dataset("PoetryMTEB/CCPD")
ex = ds["train"][0]
print(ex["poem"][:60])
print(ex["theme_label_names"], ex["emotion_label_names"])

Intended use

• PoetryMTEB multi-label probing of classical Chinese poem embeddings for theme, emotion, or joint prediction.
• Complementary single-label Tang theme set: PoetryMTEB/TCCP.

Citation

@article{wei2024ccpd,
  title   = {Knowledge-Guided Transformer for Joint Theme and Emotion Classification of Chinese Classical Poetry},
  author  = {Wei, Yuting and Hu, Lianyong and Zhu, Yang and Zhao, Jie and Wu, Bin},
  journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},
  volume  = {32},
  pages   = {4783--4794},
  year    = {2024},
  doi     = {10.1109/TASLP.2024.3487409}
}

• Dataset: https://github.com/yuting-wei/CCPD
• This Hub packaging: PoetryMTEB/CCPD

License

Distributed under CC BY 4.0 for this PoetryMTEB packaging. Please cite Wei et al. (2024) and the upstream GitHub repository.