README.md

---
license: other
task_categories:
- text-classification
task_ids:
- multi-class-classification
language:
- ta
multilinguality:
- monolingual
size_categories:
- 1K<n<10K
pretty_name: TamilSangamLiteraryDeviceClassification
tags:
- poetry
- tamil
- sangam
- literary-device
- rhetoric
- multi-class-classification
- mteb
- poetrymteb
- embedding-evaluation
annotations_creators:
- expert-generated
source_datasets:
- Tamil-Sangam-Literature-Dataset
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: line_number
    dtype: int64
  - name: poem
    dtype: string
  - name: transliteration
    dtype: string
  - name: english
    dtype: string
  - name: label
    dtype: int64
  - name: label_name
    dtype: string
  - name: original_device
    dtype: string
  splits:
  - name: train
    num_examples: 840
  - name: test
    num_examples: 211
---

# TamilSangamLiteraryDeviceClassification

Multi-class **literary device / rhetoric classification** for classical **Tamil Sangam** verse lines (PoetryMTEB), derived from the [Tamil Sangam Literature Dataset](https://github.com/kameshkanna/Tamil-Sangam-Literature-Dataset).

Upstream `Literary Devices` strings are noisy (44 free-text labels, heavily skewed to `Imagery`). We **merge** near-duplicates and rare labels into a closed **8-class** taxonomy, then create a stratified train/test split for embedding evaluation.

## Dataset Card

| Item | Description |
|------|-------------|
| **Source** | [Tamil Sangam Literature Dataset](https://github.com/kameshkanna/Tamil-Sangam-Literature-Dataset) (`Tamil Dangam.csv`) |
| **Languages** | Tamil (`ta`); English translation & Latin transliteration kept as auxiliaries |
| **Unit** | Verse **line** (`poem` = Original Tamil Text) |
| **Labels** | **8** merged literary-device classes |
| **Size** | train=840; test=211 (dropped empty text/label: 1) |
| **Splits** | Stratified by merged label ≈ 80% / 20%, seed=42 |
| **License** | Upstream repository terms (no explicit SPDX; research use — attribute source) |
| **Evaluation metrics** | Classification on embeddings: **accuracy**, **macro/micro F1** |

## Label taxonomy (8)

| id | label_name | zh | gloss | train | test | total |
|---:|------------|----|-------|------:|-----:|------:|
| 0 | `Imagery` | 意象描写 | Imagery and related visual/emotional descriptive imagery | 486 | 122 | 608 |
| 1 | `Simile` | 明喻 | Simile | 102 | 26 | 128 |
| 2 | `Descriptive` | 描述 | Descriptive / setting description | 91 | 23 | 114 |
| 3 | `Narrative` | 叙事 | Narrative / narrative reflection | 41 | 10 | 51 |
| 4 | `DirectAddress` | 直接称呼 | Direct address to a hearer / addressee | 36 | 9 | 45 |
| 5 | `Metaphor` | 隐喻 | Metaphor | 26 | 6 | 32 |
| 6 | `RhetoricalQuestion` | 反问/设问 | Rhetorical question / question | 11 | 3 | 14 |
| 7 | `Other` | 其他修辞 | Merged rare devices (symbolism, personification, hyperbole, etc.) | 47 | 12 | 59 |

Codebook: `label_taxonomy.json`.

### Merge rules (upstream → merged)

| original_device | merged label |
|-----------------|--------------|
| `Description` | `Descriptive` |
| `Descriptive` | `Descriptive` |
| `Setting` | `Descriptive` |
| `Direct Address` | `DirectAddress` |
| `Direct address` | `DirectAddress` |
| `Descriptive imagery` | `Imagery` |
| `Emotional imagery` | `Imagery` |
| `Imagery` | `Imagery` |
| `Visual imagery` | `Imagery` |
| `Metaphor` | `Metaphor` |
| `Narrative` | `Narrative` |
| `Narrative Reflection` | `Narrative` |
| `Admonition` | `Other` |
| `Conditional` | `Other` |
| `Conditional Statement` | `Other` |
| `Conditional statement` | `Other` |
| `Contrast` | `Other` |
| `Dialogue` | `Other` |
| `Direct Speech` | `Other` |
| `Direct speech` | `Other` |
| `Direct statement` | `Other` |
| `Emotional Expression` | `Other` |
| `Encouragement` | `Other` |
| `Exclamation` | `Other` |
| `Expression of Suffering` | `Other` |
| `Expression of longing` | `Other` |
| `Expression of well-wishes` | `Other` |
| `Historical Reference` | `Other` |
| `Hyperbole` | `Other` |
| `Hypothetical` | `Other` |
| `Irony` | `Other` |
| `Negative statement` | `Other` |
| `Personification` | `Other` |
| `Philosophical Reflection` | `Other` |
| `Quotation` | `Other` |
| `Reference` | `Other` |
| `Repetition` | `Other` |
| `Symbolism` | `Other` |
| `Theme` | `Other` |
| `Wish` | `Other` |
| `Question` | `RhetoricalQuestion` |
| `Rhetorical Question` | `RhetoricalQuestion` |
| `Rhetorical question` | `RhetoricalQuestion` |
| `Simile` | `Simile` |

Any upstream string not listed above is mapped to **`Other`**.

**Rationale:** `Imagery` alone covers ~57% of rows; many remaining labels have ≤4 examples and cannot support a reliable test split. Merging yields a compact, imbalanced-but-usable 8-way task.

## Features

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Example id (`sangam-device-{line_number}`) |
| `line_number` | int64 | Upstream `Line Number` |
| `poem` | string | **Classification input**: Original Tamil Text |
| `transliteration` | string | Latin-script transliteration (auxiliary) |
| `english` | string | English translation (auxiliary) |
| `label` | int64 | Merged class id (0–7) |
| `label_name` | string | Canonical English class name |
| `original_device` | string | Raw upstream `Literary Devices` string (before merge) |

## Construction method

1. Load `Tamil Dangam.csv` (1052 lines).
2. Map `Literary Devices` → 8-way taxonomy via the merge table above.
3. Drop rows with empty Tamil text or missing device label.
4. Stratified train/test split by `label_name` (seed=42, ≈ 80/19).

## How to load

```python
from datasets import load_dataset

ds = load_dataset("PoetryMTEB/TamilSangamLiteraryDeviceClassification")
print(ds["train"][0]["poem"])
print(ds["train"][0]["label_name"], ds["train"][0]["original_device"])
```

## Intended use / limitations

- PoetryMTEB multi-class probing of **Tamil poetic line** embeddings for rhetorical style.
- Labels inherit upstream annotation quality; `Imagery` remains dominant after merge.
- `Other` is a residual bucket — do not over-interpret fine-grained rhetoric inside it.
- Companion bitext release: `PoetryMTEB/TamilSangamBitextMining`.

## Citation / provenance

Please attribute the upstream dataset:

- https://github.com/kameshkanna/Tamil-Sangam-Literature-Dataset
- This Hub packaging: `PoetryMTEB/TamilSangamLiteraryDeviceClassification`

## License

Follow the upstream GitHub repository terms. This packaging redistributes line texts and merged device labels for research evaluation under PoetryMTEB.