ArWordVec is a collection of pre-trained word embedding model built from huge repository of Arabic tweets in different topics. The aim of these models is to support the community in their Arabic NLP-based research. In order to build an efficient word embedding model on Twitter Arabic data, we decided to collect millions of tweets from multiple domains and hashtags to boost theword embedding model with the diverse usage of the words within the different topics. We collected about 55 Million tweets within different topics covering the educational domain, health care, politics, social affairs, etc.

We have investigated many techniques for building word embedding models with the different parameters, which are word2vec, proposed by Google in 2013, that supports the continuous bag-of-words (CBOW) and skip-gram (SG) approaches, and GloVe toolkit, proposed by Stanford NLP Group in 2014. Several models have been generated and named as ‘model-d-w-m’. For example, CBOW-500-3-400 is the model built with CBOW approach that has vector size of 500, window size equal to 3, and the minimum word count is set to 400.

The above models were built using the gensim library that can be installed using pip or conda using the following commands:

pip install gensim

conda install gensim

After you download and extract the required ArWordVec model, for example SG-300-3-400, you can find three files named as:

1. SG-300-3-400.model
2. SG-300-3-400.model.trainables.syn1neg.npy
3. SG-300-3-400.wv.vectors.npy

You will only use the first one while your are loading the model, but the other two files must be kept in the same path.

In your python script, just add the following lines:

import gensim
file_path = 'SG-300-3-400.model'
word_embed = gensim.models.Word2Vec.load(file_path)

For more information about how we preprocess the Arabic tweets before building the models, please check preprocess_tweets.py script.

We express our thanks to the administration of the High Performance Computing Center (HPCC) at King Abdulaziz University, Jeddah, Saudi Arabia, for their support and the access to the Aziz Supercomputer that helped us in performing our experiments which require both huge computing capabilities and storage space.

M.M. Fouad, A. Mahany, N. Aljohani, R.A. Abbasi, and S.-U. Hassan, “ArWordVec: Efficient word embedding models for arabic tweets,” Soft Computing, vol. 24, no. 11, pp. 8061–8068, 2020. https://doi.org/10.1007/s00500-019-04153-6