README.md

---
license: apache-2.0
task_categories:
- text-classification
task_ids:
- multi-label-classification
language:
- es
multilinguality:
- monolingual
size_categories:
- n<1K
pretty_name: DISCOPALAffectClassification
tags:
- poetry
- spanish
- sonnet
- affect
- emotion-classification
- multi-label-classification
- disco
- disco-pal
- mteb
- poetrymteb
- embedding-evaluation
annotations_creators:
- expert-generated
source_datasets:
- DISCO PAL
- DISCO
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  default: true
dataset_info:
- config_name: default
  features:
  - name: id
    dtype: string
  - name: source_index
    dtype: int64
  - name: poem
    dtype: string
  - name: labels
    sequence: int64
  - name: label_names
    sequence: string
  - name: scores
    sequence: float64
  - name: dimensional_scores
    sequence: float64
  - name: dimensional_names
    sequence: string
  - name: n_annotators
    dtype: int64
  splits:
  - name: train
    num_examples: 219
  - name: test
    num_examples: 55
---

# DISCOPALAffectClassification

Multi-label **affect / emotion classification** over **Spanish sonnets** for PoetryMTEB, derived from [DISCO PAL](https://arxiv.org/abs/2007.04626) (Barbado et al., 2020).

Each poem is a classical **soneto** from the diachronic Spanish sonnet corpus [DISCO](https://github.com/pruizf/disco), annotated by three POSTDATA (UNED) domain experts for evoked affect on a **1–4** intensity scale. We aggregate with the official **median** file (`poems_corpus_all.csv`) and binarize the five basic emotions for embedding evaluation.

## Dataset Card

| Item | Description |
|------|-------------|
| **Source annotations** | [DISCO PAL](https://github.com/AlbertoBarbado/DISCO_PAL) (`Processed Annotations/poems_corpus_all.csv`) |
| **Source poems** | [DISCO](https://github.com/pruizf/disco) (Diachronic Spanish Sonnet Corpus) |
| **Paper** | Barbado et al., *DISCO PAL: Diachronic Spanish Sonnet Corpus with Psychological and Affective Labels*, [arXiv:2007.04626](https://arxiv.org/abs/2007.04626) |
| **Languages** | Spanish (`es`) |
| **Unit** | Full sonnet text (`poem`) |
| **Labels** | Multi-label subset of **5** basic emotions (Happiness, Sadness, Anger, Fear, Disgust) |
| **Score scale** | Continuous intensities **1–4** (evoked affect); median of 3 experts |
| **Size** | train=219; test=55 |
| **Splits** | Stratified by primary (highest-score) affect ≈ 80% / 20%, seed=42 |
| **License** | [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0) (DISCO PAL project) |
| **Evaluation metrics** | Multi-label classification on embeddings: **macro/micro F1**, **Average Precision (AP)** |

## Label binarization

Absolute thresholds on the 1–4 scale leave many poems with **no** positive emotion (e.g. score ≥ 3 empties ~45% of poems). Following the PoetryMTEB POCA pipeline, we therefore use **relative** within-poem salience:

1. Take the five emotion intensities (median over experts).
2. Compute **within-poem z-scores**; keep emotions with \(z \ge 1.0\).
3. Always include the **top-1** emotion (guarantees ≥1 label).

Mean labels/poem ≈ **1.22**.

Auxiliary **dimensional** ratings (Valence, Arousal, Concreteness, Imageability, Context Availability) are stored as continuous scores for analysis; they are **not** classification targets in this release.

## Label taxonomy (5)

| id | label_name (en) | es | zh | train | test | total |
|---:|-----------------|----|----|------:|-----:|------:|
| 0 | `Happiness` | Alegría | 喜悦 | 86 | 22 | 108 |
| 1 | `Sadness` | Tristeza | 悲伤 | 125 | 33 | 158 |
| 2 | `Anger` | Ira | 愤怒 | 14 | 4 | 18 |
| 3 | `Fear` | Miedo | 恐惧 | 10 | 2 | 12 |
| 4 | `Disgust` | Asco | 厌恶 | 30 | 8 | 38 |

Codebook: `label_taxonomy.json`.

### Dimensional scores (auxiliary)

| name (en) | es | zh | scale |
|-----------|----|----|-------|
| `Valence` | Valencia | 效价 | 1–4 (median of 3 experts) |
| `Arousal` | Activación | 唤醒度 | 1–4 (median of 3 experts) |
| `Concreteness` | Concreción | 具体性 | 1–4 (median of 3 experts) |
| `Imageability` | Imaginabilidad | 意象性 | 1–4 (median of 3 experts) |
| `ContextAvailability` | Disponibilidad contextual | 语境可得性 | 1–4 (median of 3 experts) |

## Features

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Example id (`discopal-affect-{source_index}`) |
| `source_index` | int64 | Row index in upstream `poems_corpus_all.csv` |
| `poem` | string | Full Spanish sonnet text (classification input) |
| `labels` | list[int64] | Emotion class indices |
| `label_names` | list[string] | Canonical English emotion names |
| `scores` | list[float64] | Median intensities for the 5 emotions (taxonomy order, 1–4) |
| `dimensional_scores` | list[float64] | Median dimensional ratings (order = `dimensional_names`) |
| `dimensional_names` | list[string] | Names of dimensional dimensions |
| `n_annotators` | int64 | Number of experts aggregated (3; median) |

## Construction method

1. Load `poems_corpus_all.csv` (median of annotators a1/a2/a3).
2. Normalize column names (strip whitespace / NBSP artifacts).
3. Map source columns (`happinness` typo preserved upstream → `Happiness`) to the 5-emotion taxonomy.
4. Binarize with within-poem \(z \ge 1.0\) + top-1.
5. Stratified train/test split by primary affect (seed=42).

## Intended use

Designed for **PoetryMTEB / MTEB-style** multi-label classification probing of poem embeddings in Spanish. Not a clinical diagnostic resource; psychological companion labels are released separately as `DISCOPALPsychClassification`.

## Citation

```bibtex
@article{barbado2020disco,
  title={DISCO PAL: Diachronic Spanish Sonnet Corpus with Psychological and Affective Labels},
  author={Barbado, Alberto and Fresno, Víctor and Riesco, Ángeles Manjarrés and Ros, Salvador},
  journal={arXiv preprint arXiv:2007.04626},
  year={2020}
}
```

Please also cite the underlying sonnet corpus:

```bibtex
@misc{disco2017,
  author={Ruiz Fabo, Pablo and Bermúdez Sabel, Helena and Martínez Cantón, Clara and Calvo Tello, José},
  title={Diachronic Spanish Sonnet Corpus (DISCO)},
  year={2017},
  howpublished={UNED / Zenodo},
  doi={10.5281/zenodo.1069844},
  url={https://github.com/pruizf/disco}
}
```

## License

Apache License 2.0 for DISCO PAL annotations and this redistribution. Poem texts originate from DISCO; respect upstream DISCO terms when redistributing full texts beyond research evaluation use.