Arabic-Retrieval-v1.0

🚀 Arabic-Retrieval-v1.0

This is a high-performance Arabic information retrieval built using the robust sentence-transformers framework, it delivers state-of-the-art performance and is tailored to the richness and complexity of the Arabic language.

🔑 Key Features

• 🔥 Outstanding Performance: Matches the accuracy of top-tier multilingual models like e5-multilingual-large. See evaluation
• 💡 Arabic-Focused: Designed specifically for the nuances and dialects of Arabic, ensuring more accurate and context-aware results.
• 📉 Lightweight Efficiency: Requires 25%-50% less memory, making it ideal for environments with limited resources or edge deployments.

🌍 Why This Model?

Multilingual models are powerful, but they’re often bulky and not optimized for specific languages. This model bridges that gap, offering Arabic-native capabilities without sacrificing performance or efficiency. Whether you’re working on search engines, chatbots, or large-scale NLP pipelines, this model provides a fast, accurate, and resource-efficient solution.

Model Details

Model Description

• Model Type: Sentence Transformer
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)

Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference. It is important to add the prefixes <query>: and <passage>: to your queries and passages while retrieving in the folllowing way:

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("omarelshehy/Arabic-Retrieval-v1.0")

# Query 
query = "<query>: كيف يمكن للذكاء الاصطناعي تحسين طرق التدريس التقليدية؟"

# Passages
passages = [
    "<passage>: طرق التدريس التقليدية تستفيد من الذكاء الاصطناعي عبر تحسين عملية المتابعة وتخصيص التجربة التعليمية. يقوم الذكاء الاصطناعي بتحليل بيانات الطلاب وتقديم توصيات فعالة للمعلمين حول طرق التدريس الأفضل.",
    "<passage>: تطوير التعليم الشخصي يعتمد بشكل كبير على الذكاء الاصطناعي، الذي يقوم بمتابعة تقدم الطلاب بشكل فردي. يقدم الذكاء الاصطناعي حلولاً تعليمية مخصصة لكل طالب بناءً على مستواه وأدائه.",
    "<passage>: الدقة في تقييم الطلاب تتزايد بفضل الذكاء الاصطناعي الذي يقارن النتائج مع معايير متقدمة. بالرغم من التحديات التقليدية، الذكاء الاصطناعي يوفر أدوات تحليل تتيح تقييماً أدق لأداء الطلاب."
]

# Encode query and passages 
embeddings_query = model.encode(queries)
embeddings_passages = model.encode(passages)

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings_query, embeddings_passages)

# Get best matching passage to query
best_match = passages[similarities.argmax().item()]
print(f"Best matching passage is {best_match}")

Evaluation

This model has been ealuated using 3 different datasets and the NDCG@10 metric

• Dataset 1: castorini/mr-tydi
• Dataset 2: Omartificial-Intelligence-Space/Arabic-finanical-rag-embedding-dataset
• Dataset 3: sadeem-ai/sadeem-ar-eval-retrieval-questions and is compared to other highly performant models:

Citation

BibTeX

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}