POCAAffectClassification

Multi-label affect / emotion classification for English poetry (PoetryMTEB), derived from the POCA dataset (Khan, Hopkins & Gunes, ACII 2021).

Poems are annotated on the Geneva Emotion Wheel (20 discrete affects, intensity 0–10) via Mechanical Turk; we binarize to multi-labels for embedding evaluation.

Dataset Card

Item	Description
Source	POCA supplementary data (mturk/combined.csv + poems/)
Paper	Multi-dimensional Affect in Poetry (POCA) Dataset (ACII 2021); DOI 10.1109/ACII52823.2021.9597451
Languages	English (en)
Unit	Full poem text
Labels	Multi-label subset of 20 affects
Size	train=228; test=61 (matched poems with text)
Splits	Stratified by primary (highest-mean) affect ≈ 80% / 20%, seed=42
Evaluation metrics	Multi-label classification on embeddings: macro/micro F1, Average Precision (AP)

Label binarization (from score statistics)

MTurk scores are noisy (annotator std ≈ 2.6 on a 0–10 scale) and absolute thresholds leave many empty / over-dense label sets. We therefore use:

1. Aggregate mean score per affect across annotators for each poem.
2. Compute within-poem z-scores; keep affects with (z \ge 1.0).
3. Always include the top-1 affect (guarantees ≥1 label).

Mean labels/poem ≈ 3.12.

Label taxonomy (20)

id	label_name	train	test	total
0	Admiration	56	11	67
1	Amusement	98	26	124
2	Anger	7	2	9
3	Compassion	34	4	38
4	Contempt	13	4	17
5	Disappointment	39	11	50
6	Disgust	34	9	43
7	Fear	8	1	9
8	Guilt	14	5	19
9	Hate	10	2	12
10	Interest	10	1	11
11	Joy	131	33	164
12	Pleasure	35	9	44
13	Love	40	10	50
14	Contentment	41	13	54
15	Pride	32	8	40
16	Regret	16	4	20
17	Relief	23	3	26
18	Sadness	69	16	85
19	Shame	17	4	21

Codebook: label_taxonomy.json.

Features

Field	Type	Description
id	string	Example id
title	string	Poem title
author	string	Poet
poem	string	Full poem body
labels	list[int64]	Affect class indices
label_names	list[string]	Canonical affect names
scores	list[float64]	Mean MTurk intensities (length 20, taxonomy order)
n_annotators	int64	Number of MTurk annotations aggregated

Construction method

1. Load MTurk combined.csv; group by (title, Author); average the 20 affect columns.
2. Resolve poem text from poems/ via normalized filename matching.
3. Binarize with within-poem (z \ge 1.0) + top-1.
4. Stratified train/test split by primary affect.

Citation

@article{khan_hopkins_gunes_2021,
  title={Multi-dimensional Affect in Poetry (POCA) Dataset: Acquisition, Annotation and Baseline Results},
  url={https://www.repository.cam.ac.uk/handle/1810/326293},
  DOI={10.17863/CAM.73749},
  publisher={IEEE},
  author={Khan, Akbir and Hopkins, Jack and Gunes, Hatice},
  year={2021}
}

Also: https://doi.org/10.1109/ACII52823.2021.9597451

License

Follow upstream POCA / Cambridge repository terms (research use; rights reserved by authors/publisher unless otherwise noted).