🤖 Social DeBERTa-RG: Cross-Platform Reputation Generation

Social DeBERTa-RG is a fast, aspect-aware reputation generation framework fine-tuned on social media reviews. By leveraging a **domain-adapted DeBERTa** model, it provides highly accurate reputation scores across platforms like Twitter, Yelp, and TripAdvisor.

🎯 Core Contributions

• Social DeBERTa Backbone: Pretrained DeBERTa is further tuned on millions of Twitter reviews, capturing slang, emojis, and platform-specific tone.
• Aspect-Based Reputation Scoring: The system generates fine-grained scores for aspects like delivery, cleanliness, and price.
• Cross-Platform Adaptability: Evaluation shows high transferability from Twitter to Yelp, Hotels, and Food domains.
• Real-Time Inference: Lightweight prediction layers enable high-speed aspect classification and scoring.

🧪 Methodology

1. Platform-Aware Pretraining: Social DeBERTa is trained on domain-specific tokens, hashtags, and emoji-rich reviews.
2. Aspect Term Extraction: The model detects aspect mentions using BIO tagging with contextual embeddings.
3. Polarity Classification: Each aspect is assigned a polarity label (positive, neutral, negative).
4. Score Aggregation: Final scores are calculated using aspect frequency and polarity weighting.

⚙️ Technologies

• Social DeBERTa (Twitter-trained)
• HuggingFace Transformers
• BIO Tagging for aspect extraction
• Cross-Domain Evaluation on 3+ review types

📊 Results

• +7.2% F1 improvement over base DeBERTa on Yelp aspect-polarity detection
• 0.46s average inference per review (suitable for near-real-time dashboards)
• Robust on both short (Twitter) and long (TripAdvisor) reviews

📄 Citation

@article{boumhidi2024socialdeberta,
  title={Social DeBERTa-RG: Leveraging Domain-Adaptive Pretraining for Rapid Aspect-Based Reputation Generation Across X-Platform Reviews},
  author={Boumhidi, Achraf},
  year={2024}
}