arywiki_20230101_roberta_mlm_nobots

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

Moroccan Arabic Wikipedia (aryRoBERTa_BASE)

This aryRoBERTa_BASE model has been trained from scratch on the Moroccan Arabic Wikipedia articles (after removing the bot-generated articles), downloaded on the 1st of January 2023, processed using Gensim Python library, preprocessed using tr Linux/Unix utility and CAMeLTools Python toolkit for Arabic NLP, and hosted here at SaiedAlshahrani/Moroccan_Arabic_Wikipedia_20230101_nobots. It achieves the following results on the evaluation set:

• Pseudo-Perplexity: 5,686.44

Model description

We trained this Moroccan Arabic Wikipedia Masked Language Model (aryRoBERTa_BASE) to evaluate its performance using the Fill-Mask evaluation task and the Masked Arab States Dataset (MASD) dataset and measure the impact of bot-based generation on the Moroccan Arabic Wikipedia edition.

For more details about the experiment, please read and cite our paper:

@inproceedings{alshahrani-etal-2023-performance,
    title = "{Performance Implications of Using Unrepresentative Corpora in {A}rabic Natural Language Processing}",
    author = "Alshahrani, Saied  and Alshahrani, Norah  and Dey, Soumyabrata  and Matthews, Jeanna",
    booktitle = "Proceedings of the The First Arabic Natural Language Processing Conference (ArabicNLP 2023)",
    month = December,
    year = "2023",
    address = "Singapore (Hybrid)",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.arabicnlp-1.19",
    doi = "10.18653/v1/2023.arabicnlp-1.19",
    pages = "218--231",
    abstract = "Wikipedia articles are a widely used source of training data for Natural Language Processing (NLP) research, particularly as corpora for low-resource languages like Arabic. However, it is essential to understand the extent to which these corpora reflect the representative contributions of native speakers, especially when many entries in a given language are directly translated from other languages or automatically generated through automated mechanisms. In this paper, we study the performance implications of using inorganic corpora that are not representative of native speakers and are generated through automated techniques such as bot generation or automated template-based translation. The case of the Arabic Wikipedia editions gives a unique case study of this since the Moroccan Arabic Wikipedia edition (ARY) is small but representative, the Egyptian Arabic Wikipedia edition (ARZ) is large but unrepresentative, and the Modern Standard Arabic Wikipedia edition (AR) is both large and more representative. We intrinsically evaluate the performance of two main NLP upstream tasks, namely word representation and language modeling, using word analogy evaluations and fill-mask evaluations using our two newly created datasets: Arab States Analogy Dataset (ASAD) and Masked Arab States Dataset (MASD). We demonstrate that for good NLP performance, we need both large and organic corpora; neither alone is sufficient. We show that producing large corpora through automated means can be a counter-productive, producing models that both perform worse and lack cultural richness and meaningful representation of the Arabic language and its native speakers.",
}

Intended uses & limitations

We do not recommend using this model because it was trained only on the Moroccan Arabic Wikipedia articles (after removing the bot-generated articles), <u>unless</u> you fine-tune the model on a large, organic, and representative Moroccan Arabic dataset.

Training and evaluation data

We have trained this model on the Moroccan Arabic Wikipedia articles without bot-generated articles (SaiedAlshahrani/Moroccan_Arabic_Wikipedia_20230101_nobots) without using any validation or evaluation data (only training data) due to a lack of computational power.

Training procedure

We have trained this model using the Paperspace GPU-Cloud service. We used a machine with 8 CPUs, 45GB RAM, and A6000 GPU with 48GB RAM.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 128
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-06
• lr_scheduler_type: linear
• num_epochs: 5

Training results

Evaluation results

This aryRoBERTa_BASE model has been evaluated on the Masked Arab States Dataset (SaiedAlshahrani/MASD).

Framework versions

• Datasets 2.9.0
• Tokenizers 0.12.1
• Transformers 4.24.0
• Pytorch 1.12.1+cu116