ghananlpcommunity/multilingual-emotion-ghana

🌍 Multilingual Emotion Dataset — Ghana

A multilingual emotion classification dataset covering four languages spoken in Ghana, built to support fine-tuning of small language models (e.g. SmolLM2) for emotion detection in low-resource African language contexts.

---

Overview

Languages	English, Twi, Ewe, Dagbani
Task	Emotion classification
Format	Instruction-tuned (input → structured output)
Splits	Train / Validation / Test (80 / 10 / 10)
Source	GhanaNLP Community

---

Emotion Labels

Curious · Frustrated · Insecure · Relieved · Sad

---

Dataset Structure

Each row contains four columns:

Column	Description
input	Instruction prompt with label list, language tag, and source text
output	Structured target: <emotion>Label</emotion>
language	One of: english, twi, ewe, dagbani
label	Raw label string — useful for metrics and filtering

Example Row

input

### Instruction: Classify the emotion expressed in the text below.
Reply with exactly one word wrapped in <emotion> tags.
Choose only from this list: Curious, Frustrated, Insecure, Relieved, Sad.
### Language: twi
### Text: Wɔn nyinaa seesaw, na mmaa no bi nso seesaw...
### Response:

output

<emotion>Sad</emotion>

---

Intended Use

This dataset is designed for fine-tuning small language models on multilingual African emotion classification. The structured <emotion> output format makes inference parsing reliable and unambiguous:

import re

def extract_emotion(model_output: str) -> str:
    match = re.search(r"<emotion>(.*?)</emotion>", model_output)
    return match.group(1) if match else None

---

Languages & Coverage

Language	Family	Region
English	Indo-European	National / official language
Twi (Akan)	Niger-Congo, Kwa	Ashanti, Eastern, Central regions
Ewe	Niger-Congo, Kwa	Volta region
Dagbani	Niger-Congo, Gur	Northern region

Each non-English sample is a translation of the corresponding English sample, with labels transferred by index alignment. All four language subsets are balanced — same number of rows, same label distribution.

---

Citation

If you use this dataset, please cite:

@dataset{ghananlp_emotion_2025,
  author    = {GhanaNLP Community},
  title     = {Multilingual Emotion Dataset — Ghana},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/ghananlpcommunity/multilingual-emotion-ghana}
}

---

Related Work

• GhanaNLP — NLP tools and resources for Ghanaian languages
• Pristine Twi Dataset — Clean Twi text corpus
• SmolLM2 — Target model family for fine-tuning

---

Built by the GhanaNLP Community · Advancing African language technology