FSPC — Fine-grained Sentimental Poetry Corpus

Manually labelled fine-grained sentiment classification of classical Chinese poetry for PoetryMTEB embedding evaluation.

Upstream FSPC (THU-FSPC) annotates each poem and each line into five sentiment classes (negative → positive, including implicit polarity). Source: THUNLP-AIPoet/Datasets/FSPC (mirror).

Dataset Card

Item	Description
Dataset version (PoetryMTEB)	1.0.0
Upstream version	V1.0 (5,000 poems)
Source	THUNLP-AIPoet/Datasets/FSPC
Paper	Chen et al., Sentiment-Controllable Chinese Poetry Generation, IJCAI 2019 (pdf)
Languages	Classical Chinese / Chinese (zh)
Unit	Full poem text (poem); lines originally `
Labels	Single-label 5-class holistic sentiment (label / label_name); line-level labels also provided
Size	train=4000; test=1000; total=5000
Splits	train / test only (no validation). Poem membership preserved from the previous PoetryMTEB/FSPC release (stratified by holistic sentiment, ~80/20)
License	Academic use only (THUAIPoet Datasets release note). Cite the original paper.
Evaluation metrics	Embedding classification: accuracy, macro/micro F1 (see scripts/evaluate_fspc.py)

Label taxonomy (5)

id	upstream	label_name	name_zh	name_en	train	test	total
0	1	negative	消极	negative	231	58	289
1	2	implicit_negative	隐式消极	implicit negative	1173	293	1466
2	3	neutral	中性	neutral	1062	266	1328
3	4	implicit_positive	隐式积极	implicit positive	1250	312	1562
4	5	positive	积极	positive	284	71	355

Codebook: label_taxonomy.json. Upstream field name setiments is a known typo; this packaging stores the same JSON under sentiments.

Features

Field	Type	Description
id	string	Example id (fspc-{split}-{poem_id})
poem_id	int64	1-based index in upstream FSPC_V1.0.json order
title	string	Poem title
poet	string	Poet name
dynasty	string	Dynasty
poem	string	Classification input: full poem body (newlines between lines)
label	int64	Holistic sentiment class id (0–4)
label_name	string	Holistic class name
line_labels	list[int64]	Per-line class ids (usually 4)
line_label_names	list[string]	Per-line class names
sentiments	string	Original sentiment dict as JSON (holistic, line1…line4 with upstream codes 1–5)

Schema aligns with other PoetryMTEB classification datasets (id, poem, label, label_name).

Construction method

1. Download upstream FSPC_V1.0.json (JSON Lines, 5,000 poems).
2. Map | line separators → newlines; map upstream codes 1–5 → ids 0–4.
3. Assign train/test using poem_id membership from the prior PoetryMTEB/FSPC release (no validation split).
4. Write parquet shards, dataset_infos.yaml, Dataset Card, split id lists, checksums, eval script, and annotation prompt template.

Supporting materials (this repo)

Path	Content
VERSION	PoetryMTEB packaging version
LICENSE	Academic-use-only notice
CITATION.cff / CITATION.bib	Citation metadata
metadata.json	Dataset metadata
label_taxonomy.json	Label codebook + counts
splits/train_ids.txt / splits/test_ids.txt	Split poem_id lists
checksums.sha256	SHA256 of packaged files
prompts/annotation_prompt_template.md	Reconstructed annotation guideline template
scripts/evaluate_fspc.py	Baseline probe evaluation (accuracy / F1)

How to load

from datasets import load_dataset

ds = load_dataset("PoetryMTEB/FSPC")
print(ds["train"][0]["poem"])
print(ds["train"][0]["label_name"], ds["train"][0]["label"])

Intended use

• PoetryMTEB / MTEB-style multi-class classification probing of classical Chinese poem embeddings (fine-grained sentiment).
• Research on implicit sentiment in classical Chinese poetry.

Not for commercial use without permission from the original data providers.

Citation

@inproceedings{chensentiment19,
  author    = {Huimin Chen and Xiaoyuan Yi and Maosong Sun and Cheng Yang and Wenhao Li and Zhipeng Guo},
  title     = {Sentiment-Controllable Chinese Poetry Generation},
  booktitle = {Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence (IJCAI)},
  year      = {2019},
  address   = {Macao, China}
}

Upstream: https://github.com/THUNLP-AIPoet/Datasets/tree/master/FSPC

This Hub packaging: PoetryMTEB/FSPC (version 1.0.0)

License

Redistributed for academic research only, following the THUAIPoet Datasets release note. Please cite the IJCAI 2019 paper when using the data.