Muhamm-dk/arabic-sentiment-marbert-combined

Arabic sentiment (binary), MARBERT fine-tuned on four datasets

Fine-tuned from UBC-NLP/MARBERT for negative/positive sentiment on Arabic text (MSA and dialect). Code, benchmark and error analysis: https://github.com/Muhamm-dk/arabic-sentiment-benchmark

Training data

LABR (book reviews), HARD (hotel reviews, 20,000-row training subsample, leading Booking.com score label removed), ASTD (Egyptian tweets, Negative/Positive only), AJGT (Jordanian tweets). Seed 0, 3 epochs, lr 2e-5, best epoch chosen on validation macro-F1.

Test results (macro-F1, held-out test splits, single seed)

Dataset	Macro-F1
AJGT	0.933
ASTD	0.869
HARD (no leading descriptor)	0.908
LABR	0.864

Limitations

• Binary only: no neutral class. Neutral or factual text is forced into a polarity.
• Small test sets (AJGT 180, ASTD 242): differences of a few points are within noise.
• Political tweets: labels often follow the author's stance, not wording, and the model misreads these.
• Mixed reviews ("good but...") are a frequent error source.
• Not evaluated on Gulf or Maghrebi dialect text, or on other domains.
• The four training datasets do not state licenses. Use this model for research, and check the original dataset sources before any commercial use.