Dohahemdann/deepx-arabic-absa-marbert

DeepX Arabic ABSA — MarBERT v2

Aspect-Based Sentiment Analysis model for Arabic restaurant reviews (Egyptian/Gulf/Levantine dialects + English).

Performance

Metric	Score
Joint F1	0.7831
Aspect F1 (micro)	0.8611
Sentiment F1 (micro)	0.9113

Aspects detected

food · service · price · cleanliness · delivery · ambiance · app_experience · general · none

Sentiment classes

positive · negative · neutral

Architecture

• Backbone: UBC-NLP/MARBERTv2
• Mean-pool CLS representation
• Per-aspect sentiment heads (9 independent heads)
• Hidden dim: 384 | Dropout: 0.15
• 3-stage curriculum training with pseudo-labeling

Usage

import torch
import json
from transformers import AutoTokenizer, AutoModel

# Load tokenizer + encoder
tokenizer = AutoTokenizer.from_pretrained("Dohahemdann/deepx-arabic-absa-marbert")
# Then load classification_heads.pt and model_config.json separately
# (see full inference example in the repo)

Training details

• Trained with 3-stage curriculum (head warm-up → encoder fine-tune → pseudo-label augmentation)
• Neutral-aware focal loss with boost=5.0
• Best threshold: 0.55
• Best epoch: 8