CAMeL-Lab / camelparser-dialects-EGY

CamelParser-Dialects

CamelParser-Dialects is a neural dependency parsing model for dialectal Arabic and Modern Standard Arabic (MSA), designed under the CATiB dependency formalism .

It is based on the biaffine attention parser architecture introduced by Dozat and Manning (2017), implemented using SuPar . The model leverages CamelBERT-MIX , a pretrained language model trained on a large and diverse Arabic corpus.

Full details are available in our paper: "Parsing Arabic Dialects Revisited: New Benchmarks, Models, and Insights"

📊 Model Variants and LAS (Labeled Attachment Score) on TEST

The recommended checkpoint is the all-variety model ( MSA-EGY-GLF ) , which provides the best overall cross-dialect performance.

🧠 Model Architecture

• Encoder : CamelBERT-MIX
• Parser : Deep biaffine attention (Dozat & Manning, 2017)
• Framework : SuPar
• Formalism : CATiB dependency scheme

📚 Training Data

The models are trained on combinations of the following treebanks:

• CamelTB (MSA): camel_treebank_1.1.zip
• PATB (Penn Arabic Treebank): LDC2010T13 , LDC2011T09 , LDC2010T08
• ARZTB (Egyptian Arabic Treebank): LDC2018T23
• CamelTB-Gumar (Gulf Arabic): CamelTB-Gumar.1.0.zip

🚀 Intended Use

This model is intended for:

• Dependency parsing of Arabic text
• Linguistic analysis of dialectal Arabic

🔧 Usage

For usage instructions and code, please refer to the official repository:

👉 https://github.com/CAMeL-Lab/camel_parser_dialects

📖 Citation

If you use this model, please cite:

@inproceedings{Elshabrawy:2026:camelparser-dialects,
    title = "{Parsing Arabic Dialects Revisited: New Benchmarks, Models, and Insights}",
    author = {Ahmed Elshabrawy and
              Go Inoue and
              Muhammed AbuOdeh and
              Nizar Habash} ,
    booktitle = {Proceedings of The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT)},
    year = "2026",
    address = "Palma, Spain"
}