Egyptian Arabic sentiment (CAMeLBERT fine-tuned on ASTD)

Classifies Egyptian-dialect Arabic text as positive, negative, mixed or objective. Fine-tuned from CAMeL-Lab/bert-base-arabic-camelbert-mix on the ASTD tweets, using the official 4-class split. Code, training notebook and full results: https://github.com/Younes695/egyptian-arabic-sentiment

Results on the ASTD test set (1,999 tweets)

This checkpoint is the CAMeLBERT-Mix run chosen on validation (seed 7): test macro-F1 0.533. F1 per class: positive 0.492, negative 0.542, mixed 0.321, objective 0.779.

Use

from transformers import pipeline
clf = pipeline("text-classification", model="abdElrahmanWaelYounes/egyptian-arabic-sentiment", top_k=None)
clf("الفيلم ده تحفة بجد")

For best results, clean the text the same way as in training (egysent.preprocess.clean_tweet in the GitHub repo).

Limitations

• Trained on tweets that are mostly political and from around 2013. Reviews and customer messages will look different.
• Mixed opinions and sarcasm are the weakest cases (mixed F1 about 0.33).
• ASTD is GPL-2.0; check that its terms fit your use before deploying this model commercially.