Deep Learning for Hierarchical Arabic Text Classification

Djelloul BOUCHIHA¹, Abdelghani BOUZIANE², Noureddine DOUMI³, Farouk Omar BERBOUCHI⁴, Aymen Abdelghani KIBIR⁵, Nihad MEBARKI⁶ and Badia Achouak BENAMER⁷

 

If you use a software from this page, please cite us as follows:

Djelloul BOUCHIHA, Abdelghani BOUZIANE, Noureddine DOUMI, Farouk Omar BERBOUCHI, Aymen Abdelghani KIBIR, Nihad MEBARKI and Badia Achouak BENAMER. Deep Learning for Hierarchical Arabic Text Classification. Submitted (still under review).

 

If you need help, contact Djelloul BOUCHIHA:

[email protected]; [email protected]; [email protected];

 

Before using any software from this page, please follow carefully the following notes:

 

First you have to download the WiHArD dataset from https://data.mendeley.com/datasets/kdkryh5rs2/2. Download the whole dataset as one CSV file (WiHArD.csv)

 

Note that the deep learning classifier has been implemented using the Scientific Python Development Environment (Spyder IDE, version 4.1.5):

 

Before running the classifier, you have to install some additional Python packages:

Since we are using Anaconda environment including Spyder (Python editor), then we add packages through Anaconda Prompt terminal:

 

For the first time, check the already installed packages with:

C:\...>pip list

 

 

For BERT embedding, you have to install:

For Arabic language:

C:\...>pip install -U tensorflow

C:\...>pip install -U transformers

Look at (https://pytorch.org/get-started/locally/) to install pytorch

C:\...>conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

 

 

 

Next, you will find the Python classifier source code:

ARBERT-NN_Functional_Global.py

 

 

Next are some error and warning messages that you may meet when dealing with our classifier:

 

When launching the Arabic BERT model, if you get the following error message:

RuntimeError: The size of tensor a (532) must match the size of tensor b (512) at non-singleton dimension 1

 

If you receive the following error message:

ValueError: The first argument to `Layer.call` must always be passed.