Bytte-AI/BBC_Igbo-Pidgin_Gold-Standard_NLP_Corpus

BBC Igbo–Pidgin Gold-Standard NLP Corpus (Sample)

Sample dataset: High-quality annotated data for Nigerian Igbo and Pidgin English NLP

🤗 Hugging Face • 📊 Figshare • 🌐 Website • 📧 Contact

---

📋 Overview

The BBC Igbo–Pidgin Gold-Standard NLP Corpus (Sample) is a meticulously curated collection of professionally annotated text data designed to advance natural language processing for Nigerian languages. Created by Bytte AI, this sample corpus addresses the critical scarcity of high-quality linguistic resources for African low-resource languages.

📌 Sample Dataset Notice: This is a sample dataset representing a curated subset of annotated documents. It demonstrates the annotation methodology, quality standards, and multi-task capabilities of the full corpus. Ideal for prototyping, fine-tuning, and quality benchmarking.

🎯 Key Features

• 217 annotated documents across Igbo and Pidgin English (sample size)
• 4 complementary datasets covering multiple NLP tasks
• 4,851 named entities with 7-way classification
• Human-in-the-loop annotation using Label Studio
• News domain coverage from authoritative BBC sources
• Multi-task learning ready with aligned document IDs

📊 Quick Stats

Language	Samples	Avg. Words/Doc	Tasks
Igbo	63	494	Intent, Sentiment, Segmentation
Pidgin	91	805	Intent, Sentiment, NER, Language Quality

---

🗂️ Dataset Composition

1️⃣ BBC Igbo IQS (Intent, Quality, Sentiment)

📄 bbc_igbo_IQS.csv | 219 KB | 63 samples

• Tasks: Intent classification, tone/sentiment analysis
• Labels: 4 intent types, 4 tone types
• Format: CSV with metadata and full article text
• Use Case: Document-level classification tasks

2️⃣ BBC Pidgin IQS (Intent, Quality, Sentiment)

📄 bbc_pidgin_IQS.csv | 432 KB | 91 samples

• Tasks: Intent classification, tone/sentiment analysis, language quality assessment
• Labels: 4 intent types, 4 tone types, 2 language quality types
• Format: CSV with metadata and full article text
• Special Feature: Language quality dimension captures Pidgin-English continuum

3️⃣ BBC Igbo Sentence Segmentation

📄 bbc_igbo_sentence_segmentation.json | 482 KB | 63 samples

• Task: Sentence boundary detection
• Coverage: 98.41% of documents (62/63)
• Format: Label Studio JSON export
• Use Case: Sentence tokenization, text preprocessing

4️⃣ BBC Pidgin NER (Named Entity Recognition)

📄 bbc_pidgin_NER.json | 1.4 MB | 91 samples

• Task: Named entity recognition and classification
• Entities: 4,851 annotations across 7 types
• Entity Types: PERSON, LOCATION, DATE, ORGANIZATION, EVENT, PRODUCT, MONEY
• Format: Label Studio JSON export with character offsets
• Density: 53.31 entities per document on average

---

🏷️ Label Taxonomy

Intent Classification (5 classes)

Label	Igbo	Pidgin	Description
news-reporting	38	65	Objective reporting of current events
human-interest	22	12	Personal stories and community narratives
analysis	2	6	In-depth examination of trends
opinion	1	0	Editorial viewpoints
breaking-news	0	8	Urgent, time-sensitive updates

Tone/Sentiment (4 classes)

Label	Igbo	Pidgin	Description
neutral	35	36	Balanced, objective presentation
positive	19	14	Optimistic, hopeful framing
negative	5	28	Pessimistic, concerning framing
critical	4	13	Analytical questioning or criticism

Language Quality (Pidgin Only - 2 classes)

Label	Count	Description
mixed-pidgin-english	88	Pidgin with English lexical items
pure-pidgin	3	Predominantly Pidgin vocabulary

Named Entities (Pidgin - 7 classes)

Entity Type	Count	%	Examples
PERSON	1,681	34.6%	Vladimir Putin, Donald Trump
LOCATION	1,427	29.4%	Ukraine, Moscow, Nigeria
DATE	603	12.4%	February 2022, Thursday
ORGANIZATION	593	12.2%	BBC, World Economic Forum
EVENT	244	5.0%	Russian invasion, tok-tok
PRODUCT	240	4.9%	Air Force One
MONEY	63	1.3%	Financial amounts

---

🔬 Quality Metrics

Annotation Quality

Metric	Igbo IQS	Pidgin IQS	Igbo Seg	Pidgin NER
Annotators	1	2	1	1
Avg Lead Time	291.09s	127.07s	25.50s	77.69s
Median Lead Time	23.75s	22.54s	5.51s	5.81s

High lead times indicate thoughtful, deliberate annotation rather than rushed work.

Class Balance (Shannon Entropy in bits)

Dataset	Dimension	Entropy	Interpretation
Igbo IQS	Intent	1.22	Moderate imbalance
Igbo IQS	Tone	1.54	Good balance
Pidgin IQS	Intent	1.30	Moderate imbalance
Pidgin IQS	Tone	1.87	High balance ✅
Pidgin IQS	Language	0.21	Severe imbalance*
Pidgin NER	Entities	2.31	Good balance ✅

*Reflects authentic Pidgin usage (naturally mixed with English)

Label Distribution Characteristics

Dataset	Long-tail 80% Coverage	Variance	Avg Labels/Item
Igbo IQS (Intent)	1/4 classes	313.58	2.0
Igbo IQS (Tone)	1/4 classes	211.58	2.0
Pidgin IQS (Intent)	1/4 classes	799.58	3.0
Pidgin IQS (Tone)	2/4 classes	124.92	3.0
Pidgin NER (Entities)	4/7 types	333,815.71	53.31

---

🚀 Getting Started

Installation

# Clone repository
git clone https://github.com/Bytte-AI/BBC_Igbo-Pidgin_Gold-Standard_NLP_Corpus.git
cd BBC_Igbo-Pidgin_Gold-Standard_NLP_Corpus

# Install dependencies
pip install pandas datasets

Quick Load (Hugging Face)

from datasets import load_dataset

# Load from Hugging Face Hub
dataset = load_dataset("Bytte-AI/BBC_Igbo-Pidgin_Gold-Standard_NLP_Corpus")

# Access individual splits
igbo_iqs = dataset['igbo_iqs']
pidgin_iqs = dataset['pidgin_iqs']
igbo_segmentation = dataset['igbo_segmentation']
pidgin_ner = dataset['pidgin_ner']

Load Locally (CSV/JSON)

import pandas as pd
import json

# Load IQS datasets (CSV)
igbo_iqs = pd.read_csv('bbc_igbo_IQS.csv')
pidgin_iqs = pd.read_csv('bbc_pidgin_IQS.csv')

# Load sentence segmentation (JSON)
with open('bbc_igbo_sentence_segmentation.json', 'r', encoding='utf-8') as f:
    igbo_seg = json.load(f)

# Load NER annotations (JSON)
with open('bbc_pidgin_NER.json', 'r', encoding='utf-8') as f:
    pidgin_ner = json.load(f)

print(f"Igbo IQS samples: {len(igbo_iqs)}")
print(f"Pidgin NER samples: {len(pidgin_ner)}")

Example: Intent Classification

import pandas as pd
from sklearn.model_selection import train_test_split

# Load data
df = pd.read_csv('bbc_pidgin_IQS.csv')

# Prepare features and labels
X = df['raw_text']
y = df['intent']

# Split data (stratified to maintain class distribution)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, stratify=y, random_state=42
)

# Train your model
# ... (use transformers, sklearn, etc.)

Example: Named Entity Recognition

import json

# Load NER data
with open('bbc_pidgin_NER.json', 'r', encoding='utf-8') as f:
    data = json.load(f)

# Extract entities from first annotated document
sample = data[0]
text = sample['data']['raw_text']
entities = []

if sample['annotations']:
    for annotation in sample['annotations']:
        for entity in annotation['result']:
            entities.append({
                'text': entity['value']['text'],
                'label': entity['value']['labels'][0],
                'start': entity['value']['start'],
                'end': entity['value']['end']
            })

print(f"Text: {text[:100]}...")
print(f"Found {len(entities)} entities:")
for ent in entities[:5]:
    print(f"  - {ent['text']} ({ent['label']})")

---

💡 Use Cases

✅ Recommended Applications

1. Benchmarking NLP Models

   • Evaluate pre-trained multilingual models (mBERT, XLM-R, AfroXLMR)
   • Compare performance across African languages
   • Establish baselines for future research

2. Model Training & Fine-tuning

   • Fine-tune language models on Nigerian languages
   • Train specialized NER systems for West African contexts
   • Develop intent and sentiment classifiers for news domains

3. Low-Resource NLP Research

   • Study transfer learning to African languages
   • Investigate code-switching in Pidgin English
   • Analyze morphological processing for Igbo

4. Language Technology Development

   • Build content moderation tools for Nigerian platforms
   • Create information retrieval systems
   • Support machine translation research

❌ Out-of-Scope Uses

• General-purpose models without domain adaptation (news-specific)
• Applications requiring perfectly balanced datasets
• Formal/literary language processing (colloquial news style)
• Production deployment without validation on target domain

---

⚠️ Limitations

Known Constraints

1. Sample Dataset - Limited Scale

   • This is a sample dataset (217 documents)
   • Not representative of full corpus size
   • Best for: prototyping, fine-tuning, quality demonstration
   • Not recommended for: training large models from scratch

2. Limited Training Scale

   • 63 Igbo samples, 91 Pidgin samples
   • Best used for fine-tuning, not training from scratch
   • Combine with other datasets for larger experiments

3. Domain Specificity

   • Exclusively news content from BBC
   • May not generalize to social media, technical, or conversational domains

4. Class Imbalance

   • News-reporting intent dominates (60-71%)
   • Consider class weighting during training
   • Apply oversampling for minority classes

5. Single Source

   • All content from BBC (editorial perspective)
   • Represents one variety of Igbo and Pidgin
   • Combine with diverse sources when possible

6. Temporal Coverage

   • Articles up to January 2026
   • May not reflect newer linguistic trends

Recommended Mitigations

• Data Augmentation: Use back-translation, paraphrasing
• Class Weighting: Apply inverse frequency weights
• Ensemble Methods: Combine with other African language corpora
• Domain Adaptation: Fine-tune on target domain after pre-training
• Stratified Splitting: Maintain class distribution in train/dev/test

---

📖 Data Format Specifications

IQS CSV Format

Column	Type	Description
annotation_id	int	Unique annotation identifier
annotator	str	Annotator ID
created_at	datetime	Annotation timestamp
id	str	Document ID (e.g., bbc_igbo_0001)
intent	str	Intent label
language	str	Language code (igbo or pidgin)
language_quality	str	Language quality (Pidgin only)
lead_time	float	Annotation time in seconds
raw_text	str	Full article text
source	str	Source platform (bbc)
title	str	Article headline
tone	str	Sentiment/tone label
updated_at	datetime	Last update timestamp
url	str	Original BBC article URL

Label Studio JSON Format

The sentence segmentation and NER datasets use Label Studio's export format:

{
  "id": 405,
  "data": {
    "id": "bbc_pidgin_0001",
    "raw_text": "Full article text...",
    "language": "pidgin",
    "source": "bbc",
    "url": "https://www.bbc.com/pidgin/...",
    "title": "Article headline"
  },
  "annotations": [{
    "id": 171,
    "completed_by": 3,
    "result": [{
      "value": {
        "start": 454,
        "end": 496,
        "text": "di status of Ukraine eastern Donbas region",
        "labels": ["EVENT"]
      },
      "id": "oJCzY13Q8I",
      "from_name": "entities",
      "to_name": "article",
      "type": "labels"
    }],
    "lead_time": 77.69
  }]
}

---

📚 Citation

If you use this dataset in your research, please cite:

@dataset{bytte_ai_bbc_igbo_pidgin_2026,
  author    = {Bytte AI},
  title     = {BBC Igbo–Pidgin Gold-Standard NLP Corpus (Sample)},
  year      = {2026},
  version   = {1.0},
  note      = {Sample dataset},
  publisher = {Hugging Face and Figshare},
  url       = {https://huggingface.co/datasets/Bytte-AI/BBC_Igbo-Pidgin_Gold-Standard_NLP_Corpus},
  license   = {CC-BY-4.0}
}

APA Style:

Bytte AI. (2026). BBC Igbo–Pidgin Gold-Standard NLP Corpus (Version 1.0) [Data set]. 
Hugging Face. https://huggingface.co/datasets/Bytte-AI/BBC_Igbo-Pidgin_Gold-Standard_NLP_Corpus

---

📜 License

This dataset is released under CC-BY-4.0 (Creative Commons Attribution 4.0 International).

You are free to:

• ✅ Share — copy and redistribute the material
• ✅ Adapt — remix, transform, and build upon the material
• ✅ Commercial use — use for commercial purposes

Under the following terms:

• 📌 Attribution — You must give appropriate credit to Bytte AI

See LICENSE for full details.

---

🤝 Contributing

We welcome contributions to improve and expand this corpus! Here's how you can help:

Reporting Issues

• Found annotation errors? Open an issue
• Have suggestions? Share them via [email protected]

Proposed Enhancements

• Additional annotation layers (syntax, morphology)
• Inter-annotator agreement studies
• Alignment with other African language corpora
• Expanded coverage (more articles, languages)

---

🌍 Related Resources

African Language Datasets

• JW300 - Parallel corpus including Igbo
• MENYO-20k - Yoruba-English parallel corpus
• AfriSenti - Sentiment analysis for African languages
• MasakhaNER - NER for African languages
• Lanfrica - Directory of African language resources

Tools & Models

• AfroXLMR - Multilingual model for African languages
• IgboAPI - Tools for Igbo language processing
• Label Studio - Annotation platform used for this corpus

---

📞 Contact & Support

Organization: Bytte AI
Website: https://www.bytte.xyz/
Email: [email protected]

Download Links:

• 🤗 Hugging Face: https://huggingface.co/datasets/Bytte-AI/BBC_Igbo-Pidgin_Gold-Standard_NLP_Corpus
• 📊 Figshare: https://figshare.com/articles/dataset/BBC_Igbo_Pidgin_Gold-Standard_NLP_Corpus/31249567

---

🙏 Acknowledgments

This corpus was created through the dedicated efforts of the Bytte AI annotation team. We acknowledge:

• BBC Igbo and BBC Pidgin for providing authoritative journalism in African languages
• Label Studio for the annotation platform
• The African NLP community for inspiring this work
• Our annotators for their meticulous and thoughtful work

---

📅 Version History

v1.0 (February 2026) - Initial Release

• 217 annotated samples across 4 datasets
• Tasks: Intent classification, sentiment analysis, sentence segmentation, NER
• Languages: Nigerian Igbo, Nigerian Pidgin English
• Human-in-the-loop annotation with quality metrics

---

🔮 Roadmap

While there are no immediate plans for expansion, potential future directions include:

• 🌱 Additional samples from diverse sources
• 🌱 More annotation tasks (POS tagging, dependency parsing)
• 🌱 Inter-annotator agreement studies
• 🌱 Expansion to other Nigerian languages (Yoruba, Hausa)
• 🌱 Cross-lingual alignment with English

Community feedback will help shape future development priorities.

---

By Bytte AI for African language NLP