POEMOGermanEmotionClassification

Line-level multi-label aesthetic emotion classification for German poetry (PoetryMTEB), derived from PO-EMO (Haider et al., LREC 2020).

Dataset Card

Item	Description
Source	tnhaider/poetry-emotion
Paper	PO-EMO: Conceptualization, Annotation, and Modeling of Aesthetic Emotions in German and English Poetry (LREC 2020)
Languages	German (de)
Unit	Verse line (poem field holds one line)
Labels	Up to 2 of 9 aesthetic emotions (multi-label)
train	Non-gold dual-annotator lines (union ≤2) + a poem-level subset of XMLGold moved so Nostalgia appears in train
test	Remaining XMLGold gold / adjudicated emo1/emo2
Anti-leakage	Gold poems matched in the complete TSV stay out of the dual-annotator train; no poem is split across train/test; train∩test line text = ∅
Size	train=2494; test=1070
Evaluation metrics	Multi-label classification on embeddings: macro/micro F1, Average Precision (AP)

Split anti-leakage checks

Check	Result
Gold poems matched in complete TSV & removed from dual-annotator train	47 poem_ids
Train ∩ test normalized line text	0
Train poems	113
Test poems	42
Gold poems moved to train (for Nostalgia coverage)	6 poems / 38 Nostalgia lines

Label taxonomy (9)

id	label_name	train	test	total
0	Beauty/Joy	1036	267	1303
1	Sadness	611	172	783
2	Uneasiness	610	154	764
3	Vitality	439	166	605
4	Awe/Sublime	261	154	415
5	Suspense	265	151	416
6	Annoyance	120	168	288
7	Humor	184	74	258
8	Nostalgia	38	32	70

Codebook: label_taxonomy.json.

Features

Field	Type	Description
id	string	Example id
title	string	Poem title
author	string	Poet
year	string	Publication year
period	string	Literary period (when available)
poem	string	Line text (classification input)
labels	list[int64]	Emotion class indices (0–8), length 1–2
label_names	list[string]	Canonical emotion names
source	string	dual_annotator_union / XMLGold_to_train (train) or XMLGold (test)

Construction method

1. Parse XMLGold/*.xml TEI lines; take gold attributes emo1 / emo2 (normalize to the 9-class scheme).
2. train (dual-annotator): read tsv/poemo.german.complete.prodody.emotion.tsv; union of emotion{1,2}anno{1,2}, drop NONE, cap at ≤2 (frequency-preferring). Exclude poems that fingerprint-match gold (exact or Jaccard ≥ 0.7) and any line text overlapping gold.
3. Nostalgia coverage: because non-gold train had 0 Nostalgia lines, move ~50% of Nostalgia-bearing gold poems (whole poem, seed=42) into train (source=XMLGold_to_train); the rest stay in test.
4. Drop lines with no remaining emotion labels.

Citation

@inproceedings{Haider2020,
  author = {Thomas Haider and Steffen Eger and Evgeny Kim and Roman Klinger and Winfried Menninghaus},
  title = {{PO-EMO}: Conceptualization, Annotation, and Modeling of Aesthetic Emotions in {German} and {English} Poetry},
  booktitle = {Proceedings of the 12th International Conference on Language Resources
    and Evaluation (LREC'20)},
  year = {2020},
  editor = {Nicoletta Calzolari and Khalid Choukri and Thierry Declerck and Hrafn
    Loftsson and Bente Maegaard and Joseph Mariani and Asuncion Moreno
    and Jan Odijk and Stelios Piperidis},
  address = {Marseille, France},
  month = {May},
  publisher = {European Language Resources Association (ELRA)},
  date = {11-16},
  language = {english},
  url = {http://www.romanklinger.de/publications/HaiderEgerKimKlingerMenninghaus2020LREC_PO-EMO.pdf}
}

Upstream repository: https://github.com/tnhaider/poetry-emotion

License

Follow the upstream PO-EMO repository terms for redistribution of the annotations and texts.