fr3on/egyptian-dialogue

Egyptian Arabic Dialogue Dataset

Dataset Description

This dataset contains 4,322 parallel Egyptian Arabic-English dialogue pairs with automatic domain classification. The data is extracted from TV series subtitles and features natural conversational Egyptian Arabic dialect (العامية المصرية).

Languages

• Source: Egyptian Arabic (ar_EG) - Colloquial dialect
• Target: English (en)

Dataset Summary

Egyptian Arabic is one of the most widely spoken Arabic dialects, used by over 100 million speakers. This dataset provides:

• Natural conversational dialogue
• Colloquial expressions and idioms
• Domain-classified content for specialized training
• Episode context for narrative understanding

Dataset Structure

Data Format

Each entry contains:

{
  "id": "ep01_line0001",
  "arabic": "خلاويص؟",
  "english": "Ready or not?",
  "episode": 1,
  "dialect": "egyptian",
  "language": "ar",
  "language_variant": "ar_EG",
  "genre": "dialogue",
  "domain": "general"
}

Data Fields

Field	Type	Description
id	string	Unique identifier (format: epXX_lineYYYY)
arabic	string	Egyptian Arabic text
english	string	English translation
episode	int	Episode number (for context)
dialect	string	Dialect identifier (always "egyptian")
language	string	ISO language code (always "ar")
language_variant	string	Specific variant code (always "ar_EG")
genre	string	Content genre (dialogue/narration)
domain	string	Auto-detected content domain

Dataset Statistics

Overview

• Total Entries: 4,322
• Episodes: 6
• Unique Domains: 18
• Unique Genres: 2
• Average Arabic Length: 25.9 characters
• Average English Length: 35.0 characters

Domain Distribution

Domain	Count	Percentage
general	2,143	49.6%
technology	531	12.3%
family	368	8.5%
horror	281	6.5%
medical	233	5.4%
romance	136	3.1%
weather	115	2.7%
food	104	2.4%
paranormal	86	2.0%
social	55	1.3%

Episode Distribution

Episode	Entries
Episode 1	889
Episode 2	782
Episode 3	584
Episode 4	907
Episode 5	554
Episode 6	606

Genre Distribution

• dialogue: 4,301 (99.5%)
• narration: 21 (0.5%)

Domains Explained

This dataset includes automatic domain classification using keyword-based detection:

• general - Everyday conversation without specific domain
• family - Family relationships, relatives, marriage
• horror - Scary themes, ghosts, supernatural fear
• medical - Healthcare, doctors, treatment
• technology - Computers, phones, internet, apps
• romance - Love, relationships, emotions
• paranormal - Mysterious, unexplained phenomena
• weather - Climate, meteorology, temperature
• food - Cooking, restaurants, meals
• social - Friends, gatherings, social life
• crime - Police, investigation, law enforcement
• education - Schools, universities, learning
• sports - Games, matches, tournaments
• entertainment - Movies, series, cinema
• legal - Law, court, legal matters
• news - Journalism, reports, media
• business - Companies, economy, trading
• politics - Government, elections, policy

Use Cases

✅ Recommended Use Cases

• Egyptian Arabic Translation: Train translation models specifically for Egyptian dialect
• Domain-Specific Models: Train models for specific domains (medical, legal, etc.)
• Dialect Studies: Research on Egyptian Arabic characteristics
• Conversational AI: Build chatbots for Egyptian users
• Language Modeling: Pre-train or fine-tune on Egyptian dialect
• Multi-Domain Learning: Train models aware of content domains

⚠️ Limitations

• Domain Scope: Limited to entertainment/dialogue domain content
• Register: Conversational/informal language only
• Size: 4,322 entries (relatively small for large-scale pre-training)
• Dialect Variation: Egyptian Arabic has regional sub-dialects not captured
• Context: Individual dialogue lines may lack broader narrative context

Loading the Dataset

Using Hugging Face Datasets

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("fr3on/egyptian-dialogue")

# Access the data
print(dataset['train'][0])

# Filter by domain
medical_data = dataset['train'].filter(lambda x: x['domain'] == 'medical')

# Filter by episode
episode_1 = dataset['train'].filter(lambda x: x['episode'] == 1)

Using Pandas

import pandas as pd

# Load Parquet file directly
df = pd.read_parquet("data/train-00000-of-00001.parquet")

# Analyze domains
print(df['domain'].value_counts())

# Filter and export
medical_df = df[df['domain'] == 'medical']

Training Examples

Translation Model

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, Seq2SeqTrainer

# Load dataset
dataset = load_dataset("fr3on/egyptian-dialogue")

# Load model for Arabic-English translation
model_name = "Helsinki-NLP/opus-mt-ar-en"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# Tokenize
def preprocess(examples):
    inputs = tokenizer(examples['arabic'], truncation=True, max_length=128)
    targets = tokenizer(examples['english'], truncation=True, max_length=128)
    inputs['labels'] = targets['input_ids']
    return inputs

tokenized = dataset.map(preprocess, batched=True)

# Train
trainer = Seq2SeqTrainer(
    model=model,
    train_dataset=tokenized['train'],
    eval_dataset=tokenized['test']
)
trainer.train()

Domain-Aware Training

from datasets import load_dataset

dataset = load_dataset("fr3on/egyptian-dialogue")

# Train separate models per domain
for domain in ['medical', 'legal', 'technology']:
    domain_data = dataset['train'].filter(lambda x: x['domain'] == domain)
    # Train domain-specific model
    print(f"Training {domain} model with {len(domain_data)} examples")

Data Collection & Processing

Source

• Origin: Egyptian TV series subtitles
• Language: Professional subtitle translations
• Quality: Natural, conversational Egyptian Arabic

Processing Pipeline

1. Extraction: Load from Excel subtitle files
2. Cleaning: Remove empty rows, very short entries
3. Deduplication: Hash-based duplicate removal (945 duplicates removed)
4. Domain Detection: Automatic classification using keyword matching
5. Genre Classification: Automatic dialogue vs. narration detection
6. Validation: Quality checks and statistics generation

Data Quality

• ✅ Deduplicated using MD5 hash matching
• ✅ Filtered entries < 2 characters
• ✅ Removed rows with missing translations
• ✅ Normalized whitespace
• ✅ Validated Arabic and English text pairs

Considerations for Using the Data

Egyptian Arabic Characteristics

Egyptian Arabic differs significantly from Modern Standard Arabic (MSA):

• Vocabulary: Distinct colloquial words (e.g., إزيك vs. كيف حالك)
• Grammar: Simplified structures (e.g., no case endings)
• Pronunciation: Different phonetics (e.g., ج pronounced as "g")
• Script: Informal spelling conventions in spoken contexts

Recommended Training Approaches

1. Fine-tune multilingual models rather than training from scratch
2. Combine with MSA data for better Arabic understanding
3. Use domain filtering for specialized applications
4. Consider episode context for narrative tasks
5. Balance domain distribution if training general model

Ethical Considerations

• Dialect Representation: Egyptian Arabic is one of many Arabic dialects
• Cultural Context: Translations maintain cultural nuances
• Source Attribution: Data from TV series subtitles
• Privacy: No personal information included

License

This dataset is released under the CC BY 4.0 License.

Citation

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

@dataset{egyptian_dialogue_2026,
  title={Egyptian Arabic Dialogue Dataset},
  author={fr3on},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/fr3on/egyptian-dialogue}
}

Acknowledgments

• Source: Egyptian TV series subtitles
• Processing: Automatic domain detection and classification
• Format: Parquet for efficaient loading and storage

Version History

• v1.0.0 (2025-12-17): Initial release
  • 4,322 entries
  • 18 domain categories
  • Automatic domain detection
  • Parquet format

---

Keywords: Egyptian Arabic, ar_EG, dialect, colloquial, translation, dialogue, domain classification, NLP, machine translation, Arabic dialects, conversational AI, parquet

Dataset Size: 4,322 examples | Format: Parquet | License: CC BY 4.0