Skip to content

hULMonA (حلمنا)​: tHe first Universal Language MOdel iN Arabic

Notifications You must be signed in to change notification settings

aub-mind/hULMonA

 
 

Repository files navigation

hULMonA: tHe first Universal Language MOdel iN Arabic

Paper: https://www.aclweb.org/anthology/W19-4608

Introduction

Recent state-of-the-art models in NLP (e.g., BERT, GPT, ULMFiT) utilize transfer learning by pre-training a language model on large curpos and then fine-tuning it on any downstream task. We developed the first Arabic specific universal language model, hULMonA, that can be fine-tuned for almost any Arabic text classification task. We evaluated hULMonA on Sentiment Analysis and achieved state-of-the-art on 4 Arabic datasets. hULMonA consists of three main stages:

1. General domain hULMonA pretraining

To capture the various properties of the Arabic language, we train the SOTA-ish language model AWD-LSTM on all Arabic Wikipedia.

This step is time consuming, but it should be done only once. We publish our pre-trained model, and it is availabe in models directory. To check the implementation details, or to pre-train your own LM, check build_arabic_language_model.ipynb

2. Target task hULMonA fine-tuning

The target task data (e.g., Twitter) will likely come from a different distribution than the general-domain data (Wikipedia). Therefore, fine-tuning the pretrained general-domain LM on the target task data is necessary for the LM to adapt to the new textual properties (e.g., dialects).

To fine-tune the pre-trained hULMonA on your own dataset, please check fine_tune_LM.ipynb

3. Target task classification

Finally, for downstream task classification, we augment the fine-tuned hULMonA with two fully connected layers with ReLU and Softmax activations respectively. Implementatoin details can be found here: fine_tune_LM.ipynb

How do I cite hULMonA?

Please cite this paper:

@inproceedings{eljundi2019hulmona,
  title={hULMonA: The Universal Language Model in Arabic},
  author={ElJundi, Obeida and Antoun, Wissam and El Droubi, Nour and Hajj, Hazem and El-Hajj, Wassim and Shaban, Khaled},
  booktitle={Proceedings of the Fourth Arabic Natural Language Processing Workshop},
  pages={68--77},
  year={2019}
}

Contact information

For help, issues, or personal communication related to using hULMonA, please contact Obeida ElJundi ([email protected]), Wissam Antoun ([email protected]), or Nour El Droubi ([email protected]).

About

hULMonA (حلمنا)​: tHe first Universal Language MOdel iN Arabic

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 79.9%
  • Python 20.1%