🏛️ A2NLP at StanceNakba 2026 (Subtask B)

Fine-Tuned AraBERT for Topic-Based Arabic Stance Detection

This repository contains the official implementation for the A2NLP system submitted to the StanceNakba 2026 Shared Task (Subtask B). The work was presented at the 15th International Conference on Language Resources and Evaluation (LREC'26) in Palma, Spain.

---

📝 Project Description

The A2NLP-StanceNakba2026-SubtaskB-AraBERTv02-Twitter system is designed to detect stances in Arabic social media discourse. By utilizing a prompt-based fine-tuning approach on AraBERTv02, the model identifies whether a text is Pro (مؤيد), Against (معارض), or Neutral (محايد) toward specific socio-political targets.

Technical Highlights

• Architecture: aubmindlab/bert-base-arabertv02 with a sequence classification head.
• Input Strategy: Topic-conditioned prompts using the format: الهدف: [TOPIC] [SEP] الموقف من: [SENTENCE]
• Performance: Achieved a Validation Macro-F1 of 0.8434.
• Target Domain: Dialectal Arabic (Levantine, Gulf, and Egyptian) as found on Twitter/X.

---

🚀 Getting Started

1. Installation

Clone the repository and install the necessary dependencies:

git clone https://github.com/YourUsername/A2NLP-StanceNakba2026-SubtaskB-AraBERTv02-Twitter.git
cd A2NLP-StanceNakba2026-SubtaskB-AraBERTv02-Twitter
pip install -r requirements.txt

---

🤖 Model Inference

You can use the model directly from the Hugging Face Hub:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

MODEL_NAME = "aomar85/A2NLP-STANCENAKBA2026-CROSS-TOPIC"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME)

topic = "التطبيع مع اسرائيل"
sentence = "التطبيع مرفوض بشكل كامل ولا يخدم القضية"
text = f"الهدف: {topic} [SEP] الموقف من: {sentence}"

inputs = tokenizer(text, return_tensors="pt", truncation=True, padding="max_length", max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    prediction = torch.argmax(logits, dim=1).item()

# Mapping: 0: Pro, 1: Against, 2: Neutral
print(f"Predicted Stance: {prediction}")

---

📜 Citation

If you find this work useful, please cite our paper published in the LREC 2026 proceedings:

@inproceedings{nairat2026a2nlp,
  author    = {Nairat, Alaa and Nairat, Aysar},
  title     = {A2NLP at StanceNakba Shared Task: Fine-Tuned AraBERT for Topic-Based Arabic Stance Detection},
  booktitle = {Proceedings of the 15th International Conference on Language Resources and Evaluation (LREC'26)},
  month     = {May},
  year      = {2026},
  address   = {Palma, Spain},
  publisher = {European Language Resources Association (ELRA)}
}

---

👥 Team & Affiliations

• Alaa Nairat — Birzeit University (anairat@birzeit.edu)
• Aysar Nairat — Modern University College (aysar.nairat@muc.edu.ps)

---

⚖️ License

This project is licensed under the Apache License 2.0.