ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic

Masked language models (MLM) have become an integral part of many natural language processing systems. Although multilingual MLMs have been introduced to serve many languages, these have limitations as to their capacity and the size and diversity of non-English data they are pre-trained on. In this work, we remedy these issues for Arabic by introducing two powerful deep bidirectional transformer-based models, ARBERT and MARBERT, that have superior performance to all existing models. To evaluate our models, we propose ArBench, a new benchmark for multi-dialectal Arabic language understanding. ArBench is built using 41 datasets targeting 5 different tasks/task clusters, allowing us to offer a series of standardized experiments under rich conditions. When fine-tuned on ArBench, ARBERT and MARBERT collectively achieve new SOTA with sizeable margins compared to all existing models such as mBERT, XLM-R (Base and Large), and AraBERT on 37 out of 45 classification tasks on the 41 datasets (%82.22). Our models are publicly available for research.