How it works

1. Download: the dataset (link below).
2. Build: a solid sentiment classifier (3 classes: positive / negative / neutral).
3. Submit: predictions for the test split, your notebook/code, and a short report.

Dataset

HARD – Hotel Arabic Reviews Dataset (Arabic reviews from booking platforms; includes Modern Standard Arabic and dialectal Arabic).

• Labels: positive, negative, neutral
• Fields (typical): id, review, rating/label (naming may vary per file)
• Dataset link: HARD Arabic Dataset
• Create your own train/dev/test splits and document how you did it.

What you need to do

• Preprocessing: Arabic normalization (remove diacritics; normalize alef/ya/ta marbuta; remove Tatweel), basic punctuation/URL cleanup.
• Baselines: TF-IDF (char 3–5 + word 1–2) + Logistic Regression / Linear SVM (with class weights).
• Advanced (optional): fine-tune AraBERT/ARBERT or XLM-R for 3-class classification.
• Evaluation: use Macro F1 on your dev split; report per-class P/R/F1.
• Error Analysis: highlight failure cases (sarcasm, short texts, mixed dialect).

Submission Format

• Predictions file: predictions.csv with two columns: id,label (labels in {positive, negative, neutral}).
• Code/Notebook: Jupyter notebook or repo link (data prep → training → eval → inference).
• Short Report (PDF/DOC, ≤3 pages): preprocessing, models tried, metrics, insights & error analysis, conclusion.