README.md

metadata

language:
  - ar
license: mit
task_categories:
  - text-classification
task_ids:
  - sentiment-analysis
tags:
  - arabic
  - sentiment
  - hotel-reviews
  - nlp
  - arabert
pretty_name: AraReview — Arabic Hotel Reviews Sentiment Dataset
size_categories:
  - 10K<n<100K
dataset_info:
  features:
    - name: text
      dtype: string
    - name: label
      dtype: string
    - name: rating
      dtype: int64
  splits:
    - name: train
      num_bytes: 22458814
      num_examples: 84372
    - name: validation
      num_bytes: 2807484
      num_examples: 10547
    - name: test
      num_bytes: 2807484
      num_examples: 10547
  download_size: 12762557
  dataset_size: 28073782
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

🏨 AraReview — Arabic Hotel Reviews Sentiment Dataset

A clean, balanced Arabic sentiment analysis dataset built from the HARD (Hotel Arabic Reviews Dataset). Contains 105,466 labeled Arabic hotel reviews ready for fine-tuning Arabic NLP models like AraBERT, CAMeLBERT, and AraELECTRA.

Dataset Summary

Property	Value
Language	Arabic (ar)
Size	105,466 reviews
Classes	2 (positive, negative)
Class balance	Perfectly balanced, 52,733 per class
Source	HARD Hotel Arabic Reviews Dataset
Task	Sentiment Classification

Class Distribution

Label	Count	Rating Range
positive	52,733	4–5 stars
negative	52,733	1–2 stars

Dataset Structure

Data Fields

• text — cleaned Arabic review text
• label — sentiment label (positive or negative)
• rating — original star rating (1, 2, 4, or 5)

Data Sample

text	label	rating
اسوا فندق في العالم. لايوجد اي شي مميز. الفندق قديم...	negative	1
فندق رائع وخدمه ممتازه. الموقع مميز جدا والغرف نظيفه...	positive	5
افشل فندق. فاشل بكل المقاييس	negative	1

Data Pipeline

The dataset was built with a full cleaning pipeline:

1. Label mapping

• 1–2 stars → negative
• 4–5 stars → positive
• Rating 3 not present in source data

2. Arabic text normalization

• Removed diacritics (tashkeel / حركات)
• Normalized Alef variants (إ أ آ → ا)
• Normalized Teh Marbuta (ة → ه)
• Normalized Yeh variants (ى → ي)
• Removed English characters
• Removed URLs and HTML tags
• Removed non-Arabic characters and special symbols
• Normalized whitespace

3. Balancing

• Undersampled majority class to match minority class
• 52,733 samples per class via random sampling (seed=42)
• Removed reviews shorter than 10 characters after cleaning

Source Data

Built from the HARD — Hotel Arabic Reviews Dataset:

• Original size: ~105K Arabic hotel reviews
• Source: arbml/HARD
• Domain: Arabic hotel reviews from booking platforms

Intended Use

This dataset is intended for:

• Fine-tuning Arabic language models for sentiment analysis
• Benchmarking Arabic NLP models
• Research in Arabic natural language processing

Models that work well with this dataset:

• aubmindlab/bert-base-arabertv2
• CAMeL-Lab/bert-base-arabic-camelbert-da-sentiment
• aubmindlab/araelectra-base-discriminator

How to Use

from datasets import load_dataset

dataset = load_dataset("dralsarrani/AraReview")
print(dataset)

# Fine-tuning example
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("aubmindlab/bert-base-arabertv2")

def tokenize(batch):
    return tokenizer(batch["text"], truncation=True, padding=True, max_length=128)

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

Limitations

• Domain-specific: trained on hotel reviews — may not generalize to other Arabic domains without fine-tuning
• Dialect mix: contains Modern Standard Arabic (MSA) and various Arabic dialects
• Binary only: neutral class not included due to absence of 3-star ratings in source data
• Reviews may contain mixed Arabic/English text that was removed during cleaning

Author

Danah Al-Sarrani AI Engineer