PoetryFoundationSubjectsClassification

Multi-label poetry subject / theme classification dataset for PoetryMTEB embedding evaluation, derived from Poetry Foundation poem subjects.

Dataset Card

Item	Description
Source	suayptalha/Poetry-Foundation-Poems; subject taxonomy aligned with Poetry Foundation Topics
Languages	English (en)
Size	train=10259; validation=1283; test=1283 (poems with ≥1 mapped subject)
Label type	Multi-label subjects at two granularities: L1 (coarse) and L2 (fine), with hierarchical codes
Splits	Poet-aware ~8:1:1 split (train / validation / test)
Construction	Parse original Tags → exclude occasions/holidays & emotion-like tags → alias to official subject names → assign S{i} / S{i}-{j} codes
License	GNU Affero General Public License v3.0 (same family as the upstream HF dataset)
Evaluation metrics	Multi-label classification on embeddings: macro/micro F1, Average Precision (AP); evaluate L1 and L2 separately

---

Features

Field	Type	Description
id	string	Example id
title	string	Poem title
author	string	Poet name
poem	string	Poem body
labels_l1	list[string]	Coarse subject codes, e.g. S1, S6
labels_l2	list[string]	Fine subject codes with parent index, e.g. S1-8, S6-6
labels_l1_names	list[string]	Human-readable L1 names
labels_l2_names	list[string]	Human-readable paths Parent/Child

Full codebook: label_taxonomy.json.

L1 codes

Code	Subject
S1	Love
S2	Nature
S3	Social Commentaries
S4	History & Politics
S5	Religion
S6	Living
S7	Time & Brevity
S8	Relationships
S9	Activities
S10	Arts & Sciences
S11	Mythology & Folklore

L2 codes are S{i}-{j} where j is the child index under that L1 (see taxonomy file). Example: Living/Health & Illness → S6-6 (index depends on Living child order).

---

Construction method

1. Upstream text & tags from suayptalha/Poetry-Foundation-Poems (Title, Poet, Poem, Tags).
2. Taxonomy follows Poetry Foundation Topics / Subjects, with project adjustments:
   • Occasions/holidays excluded
   • Emotion-like The Mind / Horror excluded from subject labels
   • Health & Illness placed under Living
   • Dataset aliases (e.g. Death→Death & Dying, First Love→New Love)
   • Classic Love kept as a Love L2 leaf
3. Hierarchical codes: L1=S{i}; L2=S{i}-{j} (parent index explicit in the L2 code).
4. Split: group by author, then allocate poet groups to train/validation/test ≈ 80%/10%/10%.

---

Label statistics (all splits pooled)

Level	#label types	notes
L1	11	codes in use
L2	71	codes in use
Avg L1 / poem	2.55
Avg L2 / poem	2.27

---

How to load

from datasets import load_dataset

ds = load_dataset("PoetryMTEB/PoetryFoundationSubjectsClassification")
print(ds["test"][0]["title"], ds["test"][0]["labels_l1"], ds["test"][0]["labels_l2_names"])

For embedding evaluation, encode poem (optionally prepend title), then train multi-label classifiers separately on labels_l1 and labels_l2.

---

License

Distributed under GNU Affero General Public License v3.0 (AGPL-3.0), consistent with the upstream dataset license on Hugging Face.

---

Citation / provenance

Please credit the upstream corpus and Poetry Foundation subject browsing structure:

• Dataset: https://huggingface.co/datasets/suayptalha/Poetry-Foundation-Poems
• Subjects / topics: https://www.poetryfoundation.org/topics

This Hub packaging: PoetryMTEB/PoetryFoundationSubjectsClassification.